The Intelligent Economy: A Substrate Architecture

Section 1 — The Transition

1.1 Why this document exists

The economy is being rebuilt around AI agents. Not the broader AI category. Not training, not inference, not content generation, not copilots. The narrow thing: software that perceives, decides, and acts across tools, systems, and counterparties without a human in the loop for each step. That rebuild is underway, in production, across capital markets, regulatory regimes, and the operating language of the largest enterprises on earth.

The infrastructure the economy currently runs on was built for humans clicking and typing: usernames and passwords, bank accounts and credit cards, signed contracts, identity verification done by people at banks. Agents do not fit that infrastructure. There is no clean way for an agent to prove who it is, no clean way for two agents to settle a payment between them, no clean way for a regulator to see what an agent did and why. The rebuild needs a new layer. That gap is the subject of this document.

The document is written for people who can do something about that gap. Strategy leaders at incumbent organizations. Investors deciding which categories will form and which will collapse. Governments and regulated-sector procurement bodies whose obligations are outrunning their tooling. Standards and policy participants shaping the coming decade. Founders building the operator categories that will define it.

This section establishes that the transition is real. The remaining sections trace how capital reorganizes around it, which incumbents are exposed, which new categories form, what sovereignty problem it surfaces, what substrate it requires, the path that substrate is taking, and what conversation should follow.

Most people who follow business news already feel like they know AI is enormous and changing everything, and may want to skip ahead. The specific thing this document is about is not as obvious as it looks. The new participants are software that acts on its own across systems and counterparties, autonomous and semi-autonomous physical machines including robots, self-driving vehicles, autonomous industrial equipment, and the sensor networks they coordinate through. For the sake of readability, this document uses agents as shorthand for the full set. Read the word that way wherever it appears. The evidence that the transition involving all of these participants is real is worth a few minutes. Six independent measurement layers, walked at agent-specific granularity across the two-and-a-half-year window from H2 2023 to H1 2026, point the same direction at consistent magnitude. After this section, the question is no longer open.

1.2 Six layers of evidence

Six layers: agentic capability at the frontier; production deployment inside enterprises; capital flowing into agent-specific categories; open protocols for agent coordination; regulatory codification across major jurisdictions; named-incumbent acknowledgment tied to financial targets. The layers are independent. Any single layer can be denied by isolating it. The signal cannot be denied across all six at once.

1.2.1 Capability at the frontier

METR's task-completion time horizon measures the length of a software or research task, by how long a human expert takes, at which a frontier AI agent succeeds with defined probability. The 50% horizon moved from single-digit minutes in mid-2024 to fourteen hours thirty minutes by February 2026 ¹. Post-2023 doubling time is approximately 130 days, accelerated from the seven-month doubling observed across 2019–2024 ². Exponential fit R² = 0.93.

The signal is not confined to one benchmark. SWE-bench Verified, five hundred human-validated GitHub problems released by OpenAI with the SWE-bench authors in August 2024, moved from roughly 20 percent at release to over 70 percent by late 2025 ³⁴. SWE-bench Pro, the harder enterprise-scale variant, reached approximately 46 percent on the harder split in late 2025, work that would have been single-digit a year earlier ⁵. The benchmarks the field treats as meaningful have shifted from pure reasoning (MMLU, AIME) to agentic task completion (METR-HRS, AgentBench, OSWorld, WebArena, SWE-bench Verified, SWE-bench Pro). At the start of the window, frontier agents could not autonomously complete tasks that took a human more than a few minutes. At the end of the window, they complete tasks that take a human most of a working day.

The compression is exponential, the rate of compression is accelerating, and the curve shows no sign of bending.

1.2.2 Production deployment

KPMG's AI Quarterly Pulse tracks AI-agent deployment at the organization level, at agent-specific granularity. Q2 2025: 11 percent of organizations had deployed agents in production. Q3 2025: 42 percent. Q4 2025: 54 percent ⁶. Eleven, forty-two, fifty-four across three consecutive quarters is the production migration itself.

CrewAI's 2026 State of Agentic AI surveyed 500 senior executives at organizations with $100M+ revenue and 5,000+ employees: 65 percent already using agents, 81 percent fully adopted or actively scaling, 100 percent planning to expand in 2026, 74 percent naming production deployment a critical priority ⁷. Cleanlab's August 2025 survey, restricted to organizations operating agents in production with live user interactions, captured the cohort that has crossed from experimentation to operations ⁸. Sector concentration reinforces the aggregate. Insurance, traditionally one of the slowest-modernizing regulated industries, moved from 8 percent full AI adoption in 2024 to 34 percent in 2025, the fastest curve in any major regulated sector during the window ⁹. The pattern repeats in financial services, healthcare, retail, and professional services.

The deployment is no longer experimental. It is operational.

1.2.3 Capital at agent-specific granularity

AI-general capital figures inflate the number and invite dismissal of the rest. The figures below are restricted to agent-specific evidence.

Coding agents, the category that interprets entire codebases and executes multi-step development tasks, attracted roughly $550M in 2024 and approximately $4B in 2025. Sevenfold in a single year ¹⁰. Named participants include Cursor ($2.3B raised in 2025 at a $29.3B valuation), Windsurf, and others ¹⁰¹¹. Vertical AI tripled to $3.5B in 2025 ¹⁰. Healthcare-AI produced the most dramatic individual valuations. Abridge moved to $5.3B ¹⁰. Hippocratic AI raised a $126M Series C at $3.5B ¹¹. Parallel, building web infrastructure specifically for AI agents, raised $100M Series A in November 2025 ¹¹.

Menlo Ventures tracks total enterprise AI revenue at $1.7B in 2023 expanding to $37B in 2025, twenty-two times in two years ¹². KPMG records organizations projecting average AI spending of $207M over the next twelve months as of late 2025, nearly double prior-year ¹³. Salesforce completed ten AI acquisitions in 2025, the most active acquirer in the category for the year, several explicitly targeting agent analytics, agent-mediated marketing automation, and agent orchestration ¹⁴. The agent-specific capital signal is smaller in absolute terms than the AI-general figures more commonly cited. It is also more accurate and more verifiable, and the direction is unambiguous.

This is the capital signal that survives sophisticated scrutiny.

1.2.4 Protocols

Standards keep the whole thing from collapsing under its own weight. Without them, every time two AI systems needed to work together, someone would have to build a custom integration by hand.

The Model Context Protocol was released by Anthropic on November 25, 2024 with Python and TypeScript SDKs, reference servers for Google Drive, Slack, GitHub, and databases, and roughly 100,000 SDK downloads in its first month ¹⁵¹⁶. OpenAI adopted MCP in March 2025. Google DeepMind confirmed Gemini support in April 2025. Microsoft, Amazon Web Services, Cloudflare, Bloomberg, Snowflake, and Salesforce followed across 2025 ¹⁷¹⁸. By April 2026 the SDKs recorded 97 million monthly downloads, 5,800 community-built servers, and 300 clients ¹⁹. Google announced the Agent2Agent protocol in April 2025 with more than fifty partner organizations including Salesforce, PayPal, and Atlassian ²⁰. A2A was transferred to the Linux Foundation in June 2025. In December 2025 MCP was donated to the Linux Foundation under a newly formed Agentic AI Foundation co-founded by Anthropic, Block, and OpenAI ²¹. The Universal Commerce Protocol was introduced in January 2026 for retail interoperability across discovery, checkout, and post-purchase, building on the Agent Payments Protocol, A2A, and MCP ²². Ecma International approved the Natural Language Interaction Protocol suite in December 2025 ²³.

Competing AI providers convergent on common infrastructure under vendor-neutral foundation governance, at this velocity, is an alignment that almost never occurs in computing history. The enterprise procurement question has already moved from "will there be a standard?" to "are our vendors compliant?" ¹⁷

The standards are forming faster than any comparable layer in computing history.

1.2.5 Regulatory codification

The European Union's AI Act, Regulation (EU) 2024/1689, entered into force August 1, 2024 as the first comprehensive AI regulatory framework with extraterritorial reach ²⁴. Prohibited practices and AI-literacy obligations took effect February 2, 2025. General-purpose-AI obligations took effect August 2, 2025. High-risk-system obligations apply from August 2, 2026. Full application applies from August 2, 2027. Penalty exposure reaches up to 7 percent of worldwide turnover, exceeding GDPR's 4 percent cap ²⁵. The European AI Office, within DG CONNECT, became operational February 2024 ²⁶.

The United States has followed a federated pattern. Executive Order 14179 of January 23, 2025 rescinded prior Executive Order 14110 and directed federal agencies to remove regulatory barriers ²⁷. The July 2025 America's AI Action Plan identified more than ninety federal policy actions ²⁸. A December 11, 2025 executive order moved further toward federal consolidation ²⁹. Concurrent state legislation accelerated. California enacted the Transparency in Frontier Artificial Intelligence Act in 2025. Texas passed the Responsible AI Governance Act the same year. Other states followed ³⁰. Beyond the EU and U.S., the United Kingdom continues its sector-specific approach ³¹. Japan's AI Promotion Act and Korea's AI Basic Act represent national-strategy approaches ³². Singapore's Infocomm Media Development Authority released its Agentic AI Primer in 2025, with governance principles specifically for autonomous AI ³²³³. Stanford HAI's 2026 AI Index counts 47 jurisdictions with active AI-specific legislation ³⁴. The OECD AI Policy Observatory tracks more than a thousand AI policies across more than seventy jurisdictions ³⁵.

Comprehensive AI regulation is active law in major jurisdictions, with enforcement, penalties, and phased schedules continuing through 2027. Section 5 returns to the sovereignty dimension this divergence creates.

1.2.6 Incumbent acknowledgment

Executives at large public organizations are constrained from speaking publicly about transformations of this magnitude until the transformation is operationally underway. Their statements are the most lagging of the six signals, and therefore the most confirmatory. AI was cited on 306 S&P 500 earnings calls during the September 15 to December 4, 2025 window, the highest count over a comparable window in the prior ten years, against a five-year average of 136 and a ten-year average of 86 ³⁶. Information technology and communication services each cited AI on 95 percent of Q3 2025 calls. By 2025, the language had shifted from "AI" generally to "agentic AI" specifically, and executives were tying it to financial targets and workforce composition.

Salesforce CEO Marc Benioff articulated the "Agentic Enterprise" frame on the company's Q3 2025 earnings call and adopted it as Salesforce's strategic positioning across 2025 ³⁷. Customers Bank CEO Sam Sidhu committed in April 2026 to reducing the bank's workforce from approximately 3,000 to fewer than 2,000 by 2030, tying it to a multi-year contract with OpenAI to deploy agents across commercial banking, with expected efficiency-ratio improvement from approximately 49 percent to the low 40s ³⁸³⁹. Zoom CEO Eric Yuan presented an AI clone leading the company's Q1 2026 earnings call ³⁸. Synopsys CFO Sassine Ghazi said on the company's Q2 2025 call that "agentic AI will transform engineering workflows" ⁴⁰. Oracle CEO Safra Catz said on the same window's call that, due to AI agents, she expected cloud-applications growth to accelerate over the coming year ⁴⁰. OpenAI's Sam Altman published in January 2025 that "in 2025, we may see the first AI agents 'join the workforce' and materially change the output of companies" ⁴¹. Workforce restructuring announcements citing AI rose sharply across the same window ⁴².

The most cautious voices in the economy are now naming agentic AI on the record and tying it to numbers they will be held to.

1.3 The transition is here

Six layers, agent-specific, same two-and-a-half-year window, consistent direction, consistent magnitude. Capability compressing exponentially at an accelerating doubling rate. Production deployment moving from 11 to 54 percent of large organizations across three quarters. Capital into agent-specific categories expanding sevenfold for coding agents and threefold for vertical AI in a single year. Open coordination protocols reaching vendor-neutral foundation governance within thirteen months of first release. Comprehensive AI regulation reaching active enforcement in major jurisdictions. Public executive acknowledgment shifted from "AI" to "agentic AI" and tied to named financial targets and workforce composition.

When a transition compresses this fast, the temptation is to reach for a recent precedent. SpaceX is the one most often grabbed. The reader reaching for SpaceX is making a category error. This is not one operator outpacing a slow industry. It is the operating assumptions of an entire economy changing simultaneously across six dimensions at once. No single-operator precedent prepares anyone for that.

The transition is happening. What it is being built on, and what it cannot be completed on, is the subject of everything that follows. Section 2 picks up where this section ends. The capital that already exists in the economy is beginning to reorganize around the layer this evidence demands. Domain by domain, the reorganization is already visible.

Section 2 — The New Economy

2.1 The capital is rotating, not arriving

The transition Section 1 documented is not arriving as a wave from outside the economy. It is arriving as a reorganization of capital that is already inside it. The figures already cited point at the direction. Total enterprise AI revenue moved from $1.7B in 2023 to $37B in 2025 ¹². Organizations project average AI spending of $207M over the next twelve months as of late 2025, nearly double prior-year ¹³. The single most active acquirer in the AI category in 2025 was Salesforce with ten acquisitions, several explicitly targeting agent analytics, agent-mediated marketing automation, and agent orchestration ¹⁴. These are not net-new dollars conjured by the AI cycle. They are existing IT, procurement, M&A, and operating budgets rotating from where they used to go toward where the rebuild is happening.

Three forces are driving the rotation, and each one is independently verifiable. The first is operational. Production deployment of agents inside enterprises moved from 11 percent of organizations in Q2 2025 to 54 percent in Q4 2025 ⁶. Organizations that have crossed from pilot to production have to fund the production, and the funding comes from existing budgets. The second is competitive. Named-incumbent executives are tying agentic AI to financial targets and workforce composition on the record, on earnings calls, in front of investors ^3738394041. When a peer commits publicly to an agentic operating model, every board in the same sector is forced to respond. The third is regulatory. The European Union's AI Act phases obligations through August 2027 ²⁴²⁵. The OECD AI Policy Observatory tracks more than a thousand AI policies across more than seventy jurisdictions ³⁵. The compliance work requires infrastructure investment whether or not a given organization wanted to make it.

The question is not whether capital is moving. Section 1 settled that. The question is which domains are most exposed to the reorganization, and why some workflows in those domains will move faster than others. The rest of this section walks the answer.

2.2 The diagnostic: where the human is required, the workflow holds

A useful diagnostic before walking domains. The workflows most exposed to agentic reorganization are the ones where the work is structured, the inputs are digital, the outputs are digital, and the human in the loop is there because no software could do the judgment, not because law or trust or safety requires a human signature on the action. The workflows least exposed are the ones where a human is required as a matter of regulation, fiduciary duty, clinical responsibility, public accountability, or physical embodiment. The mid-zone is the interesting one. It is the set of workflows where a human is currently required by convention or by existing infrastructure, but not by anything structural. Those are the workflows where the reorganization will happen first, because the only thing holding the human in place is software that was not designed for agents to participate.

That diagnostic is not a prediction. It is a structural lens for reading the domain-by-domain evidence that follows. The reader can apply it to their own domain knowledge after this section.

2.3 Eight domains, by exposure

Eight domains carry most of the reorganization that already shows up in verifiable evidence. They are presented in stack order from the most software-saturated downward to the most physically and constitutionally constrained.

2.3.1 Enterprise services

Enterprise services are the most exposed domain in the economy. The work is structured, digital end to end, and the human in the loop is there because no prior generation of software could do the judgment. CrewAI's 2026 State of Agentic AI, surveying 500 senior executives at organizations with $100M+ revenue and 5,000+ employees, reports 65 percent already using agents and 81 percent fully adopted or actively scaling ⁷. KPMG's Q4 2025 figure of 54 percent of organizations deploying agents in production is heavily concentrated in this domain ⁶. The named incumbents already on the record about agent deployment in their own operating models are enterprise-services vendors and their customers ³⁷.

2.3.2 Financial services

Financial services entered the transition more software-saturated than any other domain. The reorganization here is not from human operation toward agentic operation; it is from rules-based and statistical automation toward agentic operation. The named commitment in this domain is unusually concrete. Customers Bank CEO Sam Sidhu committed in April 2026 to reducing the bank's workforce from approximately 3,000 to fewer than 2,000 by 2030, tying the reduction to a multi-year contract with OpenAI to deploy agents across commercial banking operations, with expected efficiency-ratio improvement from approximately 49 percent to the low 40s ³⁸³⁹. Where a human is still required in financial services, it is typically because a regulator, a fiduciary duty, or a fraud-loss model requires a human signature, not because the work itself resists agentic execution.

2.3.3 Retail

Retail is exposed on two sides. On the operator side, named vendors are deploying agents across merchandising, marketing automation, and customer service workflows ¹⁴³⁷. On the protocol side, the Universal Commerce Protocol was introduced in January 2026 as a web standard for AI-agent interoperability with retail systems across discovery, checkout, and post-purchase, building on the Agent Payments Protocol, A2A, and MCP ²². The protocol's existence is the structural signal. It is being built because the demand for agent-to-retailer interaction at the protocol level is real. How much consumer-side agent activity is currently flowing through it is a question the next several years of data will answer; the protocol-layer evidence is what is verifiable today.

2.3.4 Healthcare

Healthcare carries a structural fact unique to the domain. The regulated labor cost is the operating cost. Documentation, prior authorization, claims adjudication, scheduling, and patient-facing administrative work consume a share of healthcare spending that no other domain matches. That share is the share most exposed to agentic reorganization, because it is the share where the human is currently required by infrastructure rather than by clinical judgment. The capital community's read on the size of that exposure is visible in valuations. Abridge, a clinical-documentation specialist, moved to a $5.3B valuation in the window ¹⁰. Hippocratic AI raised a $126M Series C at a $3.5B valuation, building healthcare AI agents specifically ¹¹. Insurance, the regulated industry adjacent to healthcare on the financing side, moved from 8 percent full AI adoption in 2024 to 34 percent in 2025, the fastest curve in any major regulated sector during the window ⁹. Where a clinician is required for clinical reasons, the workflow holds. Everywhere around the clinician, the workflow is moving.

2.3.5 Government services

Government services carry the same structural pattern as healthcare, with a different boundary. Constituent-facing administrative work, benefit eligibility, document processing, and permitting are exposed. Discretion exercised by an accountable public officer is not. The verifiable signal in this domain at the window close is procurement-side rather than deployment-side. Singapore's Infocomm Media Development Authority released its Agentic AI Primer in 2025, establishing governance principles specifically for autonomous AI systems with auditable autonomy, human oversight, and adaptive safeguards ³³. Stanford HAI's 2026 AI Index counts 47 jurisdictions with active AI-specific legislation, much of it shaping what governments themselves will procure ³⁴. Procurement guidance precedes procurement, and procurement precedes deployment. The deployment cycle in this domain is longer than the private domains, but it has started.

2.3.6 Transportation and physical operations

Transportation and the broader category of physical-operations workflows, including manufacturing and logistics, are exposed in their software-layer functions, routing, dispatch, predictive maintenance, fleet coordination, and back-office operations. Their physical-execution layers, the vehicles, equipment, and humans operating them, are constrained by embodiment in ways the other domains are not. The agentic reorganization in this domain is concentrated above the physical layer and at the coordination boundary between operators. Section 3 returns to this cohort when it walks incumbent revenue models. The relevant note for Section 2 is that the exposure here is partial by structural necessity, not by lag.

2.3.7 Utilities and energy

Utilities and energy were the slowest domain to attract agentic capital in the early window, and the reorganization here is genuinely at the front edge. The exposure is highest in grid balancing, demand response, predictive maintenance, and customer-side energy management, all of which are software-layer functions sitting on top of physical infrastructure with high sensor density. The physical infrastructure itself remains the constraint. Verifiable production-scale agent deployment in this domain at the window close is thinner than in the higher-tier domains. The trajectory is what matters: regulators in this sector have begun to require continuous monitoring and reporting at a scale humans alone cannot sustain, which is the structural condition that makes the exposure real even if the current deployment numbers are smaller.

2.3.8 Settlement

Settlement is not a domain in the same sense as the other seven. It is the layer that every agentic transaction across every other domain has to resolve into. Every payment, every contract, every record, every audit trail that holds up under regulatory and legal scrutiny eventually flows through it. The existing settlement infrastructure was built for humans authorizing transactions on behalf of humans, or institutions authorizing transactions on behalf of institutions. Agents acting on behalf of either, with chains of authorization that may pass through several agents before reaching a counterparty, do not fit that infrastructure. The verifiable signal that the settlement layer is being rebuilt is at the protocol level. The Agent Payments Protocol is foundational to the Universal Commerce Protocol's transactional layer ²². The structural payoff sits here. Sections 6 and 7 return to it.

2.4 The pattern

Eight domains look different on the surface. The pattern underneath is the same. The exposure tracks the diagnostic. Where the human in the loop is there for structural reasons, the workflow holds. Where the human is there because the existing infrastructure was built for humans, the workflow moves. The capital is rotating toward the moving share, faster in the software-saturated domains and more slowly in the embodied ones, but in the same direction in every case.

Each of the eight domains is solving the same problem locally. The solutions do not interoperate. The interoperability question, and the deeper question of who in each cohort gets to participate in the new economy and who gets reorganized around, is what Section 3 picks up.

Section 3 — Winners and Survivors

3.1 What this section does

Section 2 walked eight domains by exposure. This section walks the incumbent groups inside those domains and asks the same question at the next level of resolution. Where does each group's revenue come from today? Which part of that revenue depends on humans the existing infrastructure required, and which part depends on humans the work itself requires? Which forces pull each group toward contraction, and which path could pull it toward rebuild?

The section is a diagnostic, not a ranking. It does not predict which incumbents will survive and which will not. It surfaces the structural facts each group has to face, and lets the reader apply those facts to whichever group they know best.

Groups are presented in stack order from the deepest infrastructure layer upward to the layer that integrates everything above it.

3.2 Network infrastructure incumbents

This category splits into two structural roles: the carriers, cloud-edge operators, and wholesale connectivity providers that move bits between places, and the equipment manufacturers that design, build, and increasingly operate the physical and logical layers of enterprise networks. The two roles have different exposures.

Carriers and cloud-edge operators. Revenue is tied to traffic volume, premium routing, enterprise managed services, and the infrastructure that AI workloads themselves consume. Agent traffic is incremental traffic, and the operators are positioned to absorb it. The pressure runs upward. The value migrates when the layers above figure out how to coordinate agents at the protocol level without depending on operator-mediated managed services. The rebuild path for this group is to host the agentic protocol layers themselves rather than only transport traffic that flows over them. The donations of MCP and A2A to the Linux Foundation were partly underwritten by carriers and cloud-edge operators who recognize the layer is forming above them ²¹²⁰.

Network equipment manufacturers. A deeper structural shift is happening at the manufacturer layer. The largest equipment vendors are integrating agentic operations directly into the stacks that ship with their hardware. The pattern is now visible across the public statements, product announcements, and capability rollouts of multiple major manufacturers during the 2025–2026 window ⁴³⁴⁴⁴⁵. The agentic stack performs the work customers used to assemble through third parties: it discovers what is deployed, plans the configuration, validates proposed changes against live topology and telemetry before execution, executes remediation deterministically, and monitors, optimizes, and self-heals the environment continuously ⁴³⁴⁴.

The structural consequence is large. The work that systems integrators and professional-services firms historically captured in network deployment — discovery, design, configuration, validation, certification, ongoing operations — is being absorbed into the manufacturers' own agentic stacks. Revenue that used to flow to third-party deployment and operations practices flows back to the manufacturer.

The rebuild path here is clear in outline. Manufacturers that operate the agentic layer on top of their own equipment capture the integration and operations economics that used to sit above them. Manufacturers that ship equipment without the agentic layer become commodity providers underneath someone else's stack. The same diagnostic applies in adjacent equipment categories: storage, compute, security appliances, and the silicon underneath them. What this rebuild surfaces is a requirement no single manufacturer can satisfy on its own: cross-vendor agent coordination, auditability of automated design and operations decisions, and accountability that travels with the system through every change. Each manufacturer's stack is a silo today. The interoperability layer underneath them is not yet built.

3.3 Enterprise productivity-and-cloud incumbents

These are the large platforms that sell productivity software, collaboration tools, and the cloud capacity underneath them. Revenue is per-seat, per-workload, per-tenant, and increasingly per-agent.

This group is the most directly named in the agentic transition's public language. The named-incumbent commitments in Section 1 are concentrated here ³⁷¹⁴. The diagnostic is unusually clean. The seats this group sells are the seats the customer's agentic deployments are replacing or restructuring. The defense is that the agentic layer attaches to the same productivity surface the seats already cover, so the per-seat model becomes a per-agent or per-workflow model rather than disappearing outright. The rebuild path is what this group is already executing in public, which is why the language has moved from "AI features" to "Agentic Enterprise" inside the group's own marketing ³⁷. The question facing each incumbent is whether the agentic layer they sell sits inside their own platform or sits underneath a different layer that customers reach through somebody else's gateway.

Inside this group, the collaboration-platforms subset sits in a structurally distinct position. The work running over collaboration platforms is human-to-human by design. Agents enter that work as augmentation: meeting summaries, action capture, workflow handoff to downstream systems, support of human decision cycles, rather than as replacement of the work the platform exists to carry. The defense is the same as the rest of this group, and the rebuild path is the same. The diagnostic just lands differently, because the seats sold here are seats for human collaboration that agents are now attached to, not seats that agents are taking over.

3.4 Retail commerce platforms and marketplaces

These are the platforms that host catalogs, payment rails, fulfillment networks, advertising surfaces, and the storefronts customers transact through. Revenue is take-rate on transactions, advertising spend on the platform, fulfillment fees, and cloud capacity sold to other commerce operators.

The structural mismatch here is large and underdiscussed. The catalogs, product images, descriptions, ratings, reviews, recommendation surfaces, search interfaces, and checkout flows the current platforms have spent two decades optimizing are all optimized for humans. They are not optimized for agents. Product images compress information humans extract visually that agents would extract from structured attributes. Reviews compress signal humans extract from narrative that agents would extract from structured outcomes data. Search interfaces are tuned for human query patterns and click-through behavior. Checkout flows assume human authentication and human authorization at the final step. Every layer of the current retail experience is built for human cognition operating through a screen.

Not every retail experience will move to agents. Where the purchase depends on the buyer's own sensory experience, personal judgment, lifestyle alignment, or the emotional weight of high-consideration spending, the human stays at the center of the transaction. Augmentation extends those experiences rather than removing the human from them. In purchases where the buyer's specifications are stable, the comparison is between known attributes, the cadence is recurring, and the value of human attention is low, the human steps out of the loop and an agent transacts on the human's behalf. Most retail volume sits in the second category, not the first.

The biggest near-term shift is in organization-to-organization procurement. It is the most natural category to automate first. The requirements are structured, the catalogs are already partially machine-readable, the approval chains are already digitized, and the transaction sizes justify the integration cost. Once organization-to-organization commerce runs through agents at scale, the consumer side follows along the same protocol and identity rails that the organizational side establishes. The consumer side will operate at a larger transaction count, smaller average ticket, and substantially greater aggregate scale than the organizational side that precedes it.

The Universal Commerce Protocol's introduction in January 2026 was the structural signal that agent-to-retailer interaction is being built as a web standard rather than a per-platform integration, building on the Agent Payments Protocol, A2A, and MCP ²². A platform whose moat depended on being the only path a buyer could reach a seller through has a different moat in a world where agents discover and transact across platforms on a shared protocol. The rebuild path is to be the participant in the protocol layer that the most agent traffic prefers, on the basis of selection, pricing, fulfillment reliability, returns experience, and trust anchoring rather than on the basis of being the only available surface. What this rebuild surfaces is a set of requirements the existing protocols only partially address. Buyers, sellers, agents acting on behalf of either, and the financial counterparties standing behind every transaction all need to be verifiable at the same layer. Discovery and negotiation between agents need to be governed, not ad hoc. Payments need to carry their own provenance into the ledger. Platform policy, jurisdiction-specific consumer protection, and tax remittance need to bind to each transaction at the moment it happens rather than after the fact. None of these requirements has a complete solution in production today.

3.5 Financial services and payments incumbents

These are the banks, payment networks, brokers, insurers, and exchanges that carry money and risk between counterparties. Revenue is interchange, net interest margin, asset management fees, premium income, transaction fees, and float.

The named commitment in Section 1 sits in this group ³⁸³⁹. The diagnostic is sharp. A meaningful share of this group's cost base is humans the existing infrastructure required, not humans the regulators require. Branch staffing, back-office processing, claims operations, onboarding, collections, mid-office reconciliation, much of compliance reporting — all sit in the first category. Where regulators and fiduciary duty require a human signature, the workflow holds. Everywhere else, the cost base is exposed. This industry already operates as the trust anchor for human counterparty identity. The rebuild surfaces what it does not yet anchor. Agents transacting in regulated flows on behalf of humans, organizations, and counterparties need an identity layer that fits inside existing financial regulation. Agent-to-agent transactions need cryptographic provenance that holds up under the same audit standards as human-originated transactions. Negotiation, clearing, and settlement between agents from different organizations need governed coordination rather than per-counterparty integration. Jurisdictional compliance, fraud controls, and fiduciary duty need to bind to every transaction at machine speed, not at end-of-day batch. The rebuild path runs through whichever institutions can anchor what the new economy requires the same way they anchor what the existing economy requires.

Insurance, inside this group, lands on its own structural surface. Insurers run decision-intensive workflows: underwriting, claims adjudication, coverage determination, dispute resolution. These determine outcomes for the humans the insurers cover. Agentic systems are entering these workflows fast. The accountability the insurer owes to regulators, ombudsmen, courts, and policyholders does not loosen when an agentic system enters the decision path; it tightens. Each decision has to be defensible against challenge, with the criteria, the inputs, and the reasoning available in a form a regulator can audit and a policyholder can contest. The existing infrastructure was built to produce that evidence for humans deciding; it was not built to produce it for agentic systems deciding at machine speed and at agentic-volume scale. The rebuild surfaces an audit-trail requirement that anchors every decision to identity, captures the policy and inputs the decision rested on, and stands up to whoever later asks for it.

3.6 Healthcare incumbents

These are the providers, payers, pharmacy benefit managers, and clinical-software vendors that operate the healthcare system. Revenue is fee-for-service, capitation, premium income, manufacturer rebates, and software licensing. The industry sits inside a constituent-mandated environment. Patients, providers, and regulators each impose obligations that cannot be avoided, which means the industry cannot simply rebuild around an agentic layer the way a fully private group can. The rebuild has to be negotiated with the constituents, in their language and on their terms.

The administrative-layer exposure is large and was framed in Section 2 at domain level ¹⁰¹¹. The clinical-layer exposure is more advanced than is commonly recognized. The U.S. Food and Drug Administration had authorized 1,451 AI-enabled medical devices through the end of 2025, of which 1,104 were radiology devices, representing approximately 76 percent of the total list ⁴⁶. Named manufacturers include GE HealthCare (120 authorized radiology AI products, including acquisitions Bay Labs, BK Medical, Caption Health, MIM Software, icometrix, and Spectronic Medical), Siemens Healthineers (89), Philips (50), Canon (45), United Imaging (38), Aidoc (31), and DeepHealth (28) ⁴⁶. The most clinically substantive deployments include image triage that prioritizes urgent cases ahead of routine review, image-interpretation assistance for nodule detection and fracture detection, stroke and pulmonary-embolism detection in real time, and increasingly preliminary report drafting ⁴⁷. AI-enabled medical imaging is projected to grow from $1.67B in 2025 to $26.23B by 2034 at a compound annual growth rate exceeding 35 percent ⁴⁷. Patient-intake workflows, claims adjudication, prior authorization, and parts of preliminary diagnostic reasoning are progressing on the same trajectory.

The diagnostic still applies. Where a clinician's judgment, license, and accountability are required for clinical reasons, the workflow holds. Where the existing infrastructure required a human in front of a screen because no prior software could do the work, the workflow is moving. The rebuild surfaces requirements healthcare's existing infrastructure does not address. Patients, providers, payers, and the agents operating on behalf of any of them need to be tied into a single verifiable flow. Clinical-system agents, administrative-system agents, and patient-facing agents need governed coordination across organizational boundaries within HIPAA and equivalent regimes. Claims, payments, and reimbursements need provenance that holds up to regulators and auditors. Clinical policy, payer policy, and patient consent need to bind to every interaction in real time. The rebuild path is for incumbents to absorb these requirements into their existing constituent-facing obligations rather than have the agentic layer built around them by a non-incumbent operator that would have to recreate the regulatory relationships from zero.

3.7 Physical-operations incumbents

These are the construction, manufacturing, agriculture, logistics, and engineering-and-architecture incumbents whose work product is physical and whose revenue depends on coordinating skilled labor, equipment, materials, and customer-side stakeholders to deliver it.

This group's exposure includes a category the other groups do not face directly. The physical machines and the sensor networks they depend on are themselves participants in the agentic layer. Autonomous and semi-autonomous vehicles, autonomous agricultural and construction equipment, smart buildings, and industrial sensor networks all have to be addressable, identifiable, and accountable in the same identity layer as the human and organizational participants. They are first-class participants in the new economy, not peripherals attached to it. The diagnostic runs through the coordination layer above the physical work. Where humans coordinate physically, the work holds. Where humans coordinate through software that could let agents and machines coordinate directly, the work moves. The rebuild path is for incumbents in this group to operate the coordination layer between physical participants in their own existing customer relationships, rather than have it operated by a non-incumbent layer above. What this rebuild surfaces is a set of requirements no current infrastructure satisfies at the necessary scope. Every participant, human, organizational, and physical, needs to be addressable at the same naming surface. Vehicles, equipment, sensors, and software agents need governed coordination across operator boundaries. The per-asset, per-job, per-shipment, and per-cycle transactions physical operations generate at high volume need provenance and settlement at machine speed. Safety, environmental, labor, and jurisdictional policy need to bind to every coordinated action. None of this exists end to end today.

3.8 Government services and constituent-facing public-sector operators

These are the agencies, public-sector vendors, and constituent-facing service operators that deliver services on behalf of governments. Revenue is appropriations, contracts, and fee-funded service streams.

This group sits inside the deepest constituent mandate in the document. Citizens, regulators within other branches of the same government, and oversight bodies all impose obligations that cannot be avoided. The rebuild has to satisfy a higher standard of auditability and accountability than any private group faces.

The exposure is also unusually large in absolute terms. Constituent-facing administrative services consume a meaningful share of every major government's operating budget, and they consume it in workflows that are highly visible to constituents — passport and identity-document renewal, driver's-license issuance and renewal, tax filing and processing, benefits administration and eligibility determination, permitting across construction, business, environmental, and licensing categories, public-records requests, court filings, and the support layers that surround all of the above. Each of these workflows is structured, document-heavy, rules-driven, and currently funded by humans the existing infrastructure required, not by humans the underlying obligation requires. Each of them is also a recurring source of constituent frustration about wait times, error rates, and inconsistent service. The political dynamic is not the usual one of efficiency advocated by reformers against entrenched defenders. It runs the other direction: constituents themselves are increasingly the ones demanding the rebuild, because they experience the current system as inefficient and opaque.

A second exposure runs alongside the first. Government services are also the workflows where the use of agentic infrastructure could most directly increase public trust if implemented well. Presenting how a benefit decision was made, what evidence was considered, how a permit was reviewed, how taxpayer money moved from appropriation to outcome — all of these become possible at scale when the underlying workflows are agent-mediated and auditable by design. A government that can show its constituents exactly where their money goes, exactly how decisions were made, and exactly what was processed in their name has a political asset previous governments did not have. The rebuild here is not only a cost-cutting story. It is a transparency story, and transparency is something constituents reward at the ballot box.

The verifiable signal at the window close is procurement-side rather than deployment-side. Singapore's Infocomm Media Development Authority released its Agentic AI Primer in 2025, establishing governance principles specifically for autonomous AI systems with auditable autonomy, human oversight, and adaptive safeguards ³³. Stanford HAI's 2026 AI Index counts 47 jurisdictions with active AI-specific legislation, much of it shaping what governments themselves will procure ³⁴. Procurement guidance precedes procurement, and procurement precedes deployment.

The diagnostic in this group is sharp. Discretion exercised by an accountable public officer holds. Constituent-facing administrative work, eligibility determination, document processing, permitting, identity renewal, and tax processing are exposed. The rebuild surfaces sovereignty-anchored requirements that Section 5 returns to. Identity issued or anchored by the government itself, replacing or augmenting physical credentials, is one of the largest unrealized opportunities in this group. Cooperation between citizen-facing agents and government-operated agents needs to be governed by the government's own policy, not by a private operator's. Settlement of benefits, tax remittance, fees, and inter-agency transfers needs to flow with the provenance public auditors require. Statutory authority, procedural fairness, and constituent rights need to bind to every interaction. The rebuild path is for governments to operate the agentic layer in front of their own constituents, on their own authority, rather than have a non-government operator stand between the government and the people it serves.

3.9 Services and integration providers

These are the consultancies, systems integrators, managed-services providers, and outsourcing operators whose revenue depends on standing between technology vendors and the customers deploying their technology. Revenue is hourly billing, retainer engagements, managed-services contracts, and outsourcing of operational functions.

This group sits at the top of the stack and integrates across every other group. The exposure is structural in a particular way. Each of the groups above is rebuilding its operating model around agents, which compresses both the scope and the duration of the work this group is paid to do on top of those operating models. The pressure is compounded by what is happening inside the manufacturer category itself. As described in 3.2, network equipment manufacturers are absorbing discovery, design, configuration, validation, certification, and ongoing operations into their own agentic stacks. The same pattern is emerging in other equipment categories. Capital that used to flow to third-party integration and managed-services practices is being repatriated upstream to the manufacturers running the agentic deployment stacks.

At the same time, the rebuild itself is creating short-cycle, high-skill work this group is uniquely positioned to deliver. The diagnostic is whether the revenue depends on the long-cycle deployment and operations work that compresses or on the short-cycle rebuild work that expands. The rebuild path for this group is to position into the rebuild itself, with practices, capabilities, and partnerships that operate at the agentic layer rather than at the layer it is replacing. The firms that move first into agent-layer design, identity-layer architecture, governance-layer implementation, and the customer-side configuration of manufacturer-operated agentic stacks capture the new work. The firms that hold position in the old layer compress along with it.

3.10 The pattern across all of them

Nine readings — eight industry groups and one stack-spanning group — different revenue models, different constituent obligations, different exposures. The pattern across them is the same as the pattern across domains. Where the human in the loop is there for structural reasons (regulatory, fiduciary, clinical, public-accountability, or physical-embodiment), the work holds. Where the human is there because the existing infrastructure required a human, the work is exposed. Every group has both kinds of work in its revenue base, in different proportions. The rebuild path in every case runs through the same structural requirement: a layer underneath that lets humans, organizations, governments, physical machines, and software agents all participate as first-class identifiable, accountable counterparties.

These are the groups the rebuild touches first because each one carries a critical mass of structured, software-mediated work. The diagnostic generalizes beyond them. Any industry whose revenue depends on workflows where the human is required by infrastructure rather than by structure has the same exposure. The reader who works inside or alongside an industry not named in this section can apply the diagnostic and arrive at the answer for themselves.

One structural surface unites many of those industries and many of the ones not walked: workflows that produce decisions about humans. Claims adjudication in insurance, eligibility determination in government, hiring screening in services, application processing in immigration regimes, customer-impact decisions in any consumer-facing industry. Wherever those workflows exist, the accountability for the decisions does not loosen when an agentic system enters the path; it tightens. The decision has to be defensible under regulatory scrutiny, under courts, and under the people the decision was made about. The agentic regime generates these decisions at a volume and speed the existing accountability infrastructure was not built to follow. The rebuild surfaces an audit-trail requirement that the substrate has to carry, in a form readable by regulators, courts, and the principals themselves.

What gets built underneath all of this, and who builds it, is the subject of the next sections. Section 4 walks the categories new operators are building from zero. Section 5 raises the sovereignty problem that universal participant identity surfaces.

Section 4 — The Opportunity for Builders

Section 3 ended on a structural requirement that none of the incumbent groups can satisfy on its own: a layer underneath that lets humans, organizations, governments, physical machines, and software agents all participate as first-class identifiable, accountable counterparties. The incumbents need that layer. They are not the ones who will build most of it.

That is the pattern every infrastructure transition follows. When a new substrate forms underneath an economy, the incumbents adapt their existing businesses to sit on top of it, and a new generation of operators is born to build and run the substrate itself. The companies that built the dominant businesses of the web era were not, for the most part, the companies that had dominated the era before it. The certificate authorities that made e-commerce safe to transact through did not exist before there was a web to secure. The content delivery networks that made the web fast did not exist before there was traffic to accelerate. The substrate created the categories, and the categories were filled by operators who started from zero because there was nothing there before them.

The agentic transition is opening the same kind of greenfield. This section walks the categories where new operators are building into the gap the previous two sections identified. A reader who follows the market closely will recognize that early entrants are already moving in several of these categories. That recognition is the point. The early activity confirms the requirement is real. What it has not yet produced is anything shared, interoperable, or agreed. Each entrant builds the piece its own business needs, inside its own system, to carry its own product forward. There is no agreement on what the categories even are, what they should provide, or what standard should govern them. The list below is a reading of what has to exist for the rebuild to complete, and therefore of where the building is going to happen.

4.1 Universal identity infrastructure

The first thing the rebuild requires is also the most foundational. Before any participant can transact, coordinate, or be held accountable, it has to be identifiable, and so does every other party it interacts with. Today there is no shared way to answer the question "who or what is this." A human has a passport, a driver's license, a set of logins. An organization has a registration number in one jurisdiction and a tax identity in another. A government agency has whatever credential its own systems issue. A physical machine that operates autonomously, a delivery robot, an autonomous vehicle, a piece of agricultural equipment, a sensor network inside a building, has, in most cases, nothing at all that travels with it across systems. A software agent has, at best, a key that belongs to whoever deployed it.

Some of the largest technology vendors have announced identity systems for agents, and the announcements themselves are a signal that the requirement is real. But each one issues identity inside its own platform, recognized by its own systems, governed by its own rules. None of them interoperate. None of them carry across the boundary between one operator's system and another's, and none of them extend to the full set of participants the new economy contains. The rebuild requires an identity layer where every kind of participant, human, organizational, governmental, machine, and software, can hold an identity that is recognized everywhere it needs to act, and that binds the participant to what it does. This is the deepest greenfield in the new economy, and it is the one everything else rests on.

4.2 Capability certification authorities

Once a participant can be identified, the next question is what it is allowed and able to do. A human professional carries credentials that attest to verified capability: a professional license to practice medicine, to fly a commercial aircraft, to audit financial statements. The credential is issued by a body independent of the practitioner, and the party relying on it does not have to take the practitioner's word.

Nothing equivalent exists for the new participants. A software agent deployed to act on a person's or an organization's behalf carries no independent attestation of what it was built to do, what it has been tested against, or where the boundary of its competence sits. The same gap applies well beyond software agents. The large models that agents are built on, the robots that operate in warehouses and on roads, the autonomous vehicles that carry people, the sensor networks that feed decisions, all of them act in contexts where the cost of unverified capability is high, and none of them carry a portable, independent certification of what they are safe to do. The rebuild requires operators who certify the capability of every kind of participant the way professional bodies and testing laboratories certify human and equipment capability. This is a category with direct lineage. The organizations that test and certify electrical equipment, that rate the crash safety of vehicles, that audit financial controls, all built durable businesses on the trust their attestation carries. The agentic equivalent does not exist yet, and the demand for it is created the moment these participants start acting where the cost of being wrong is real.

4.3 Reputation aggregation and publication

Certification attests to what a participant was built and tested to do. Reputation records what it has actually done. The two are different, and a functioning economy needs both. Credit bureaus, rating agencies, and review platforms all occupy the reputation role in the human economy, and each one built a business on aggregating dispersed records of past behavior into a signal others can rely on before they transact.

Early versions of agent reputation are beginning to appear, mostly as ratings inside a single platform's own marketplace. They confirm the need. They do not meet it. A rating that lives inside one platform does not travel when the participant acts somewhere else, cannot be checked by a counterparty on a different system, and in many cases can be reset by transferring ownership. The rebuild requires operators who build and publish a record that is portable, tamper-evident, and bound to the participant's identity rather than to the platform it happened to act on. The demand grows in direct proportion to the volume of agent-to-agent transactions, because the more participants transact with parties they have never encountered before, the more they need a way to know who they are dealing with.

4.4 Settlement infrastructure for agent-to-agent transactions

When two agents transact, value has to move, and the movement has to be final, recorded, and provable after the fact. Human and institutional commerce settles through banking rails, card networks, clearing houses, and the correspondent system underneath them. It also still settles, in a great deal of the world, in cash. And for most of human history it settled in whatever the parties agreed carried value, gold, grain, cattle, and at various times and places, seashells. The common thread across every one of these, from the cowrie shell to the wire transfer, is a human deciding to hand something over. An agent can do none of it. It cannot pass cash across a counter, and it certainly cannot pass a seashell. The only value an agent can move is value that has been rebuilt as something an agent can hold, send, and prove it sent, and almost none of the apparatus for that exists yet.

Around settlement sits an entire apparatus the agentic economy does not have. Humans have contracts that define what was agreed. They have courts that interpret those contracts when the parties disagree. They have dispute rails, chargebacks, arbitration, and escrow held by a trusted third party, that resolve the cases where something goes wrong. Agents have none of this. They can move a payment, once the value is in a form they can move. They cannot, today, hold funds in escrow against delivery, release a partial payment when a milestone is met, attach a warranty to a result, or resolve a dispute when a counterparty fails to perform.

Two facts about this layer matter more than the rest. The first is that a payment is the smallest part of a transaction. The hard parts are escrow against delivery, milestone-based partial payment, warranties on what was delivered, and dispute resolution when the delivery falls short, and none of these exist for agents in any shared form. The second is scale, and it is the fact that forces the architecture. The largest card network on earth processed about 234 billion transactions in its most recent fiscal year, on the order of hundreds of millions a day, and the largest interbank messaging network carries on the order of tens of millions of messages a day ⁴⁸⁴⁹. Those volumes are bounded by human pacing, because humans originate the transactions at the cardholder side, even when authorization and clearing run end-to-end without a human in the loop. Agent-to-agent transactions are bounded by nothing of the kind. They execute continuously, in fractions, across vast numbers of simultaneous interactions, at machine speed. The volume an agentic economy generates will exceed the volume of the human-paced systems by orders of magnitude, and it will do so precisely because there is no human at the decisive step to slow it down. That is also why there cannot be one. No population of human reviewers could approve the transaction volume an agentic economy produces. An economy that required a human to review and approve every agent transaction, every escrow release, every warranty claim, every dispute, would not run slowly. It would not run at all.

The protocol activity already names this need: an Agent Payments Protocol and a Universal Commerce Protocol have been introduced to handle agent-mediated transactions ²²²³. Protocols define how the messages move. The rebuild also requires operators who run the settlement and the apparatus around it, escrow, milestones, warranties, and dispute resolution, with the provenance and auditability agent commerce will be held to, and with no human in the loop at the moment of execution. And because these operators and arbitrators sit in the position of trust, deciding when value moves and who is right when a deal goes wrong, they cannot be the ones who vouch for themselves. They in turn have to be certified by an independent authority, which points back to the certification gap above and is the first sign that these categories are not a list of separate businesses but a set of parts that depend on one another.

4.5 Negotiation infrastructure

Before two parties settle, they have to agree on terms, and in the human economy that agreement is reached through negotiation. A buyer knows the price they will pay and the point they will walk away, the seller knows the lowest number they will accept, and neither reveals it. Negotiation works precisely because the information stays asymmetric. Agents cannot do this today, because nothing in the open market lets an agent hold a concealed limit while still finding common ground with a counterparty whose limit is equally concealed.

Picture the near future this points at. A company's procurement agent negotiating supply terms with a supplier's agent. A household's assistant agent negotiating the utility rate, the car-repair estimate, the grocery order across competing stores. For any of this to happen at scale, every party on the other side, every retailer, every service provider, every utility, has to be able to negotiate agent-to-agent while protecting its own position. Early forms of agent negotiation are already in use, where an agent negotiates against a human-run counterparty, and the new commerce protocols gesture at negotiation as a stage in the buying journey ⁵⁰. But genuine agent-to-agent negotiation, with each side's position structurally protected, has not yet reached working interoperable infrastructure that the market can rely on ⁵¹, and the gap remains one of the largest unbuilt functions in the entire transition.

4.6 Governance and policy enforcement

The foundational categories above, identity, certification, reputation, settlement, and negotiation, make agent activity possible. They do not, on their own, make it governable. An organization deploying agents into its operations, or allowing external agents to act against its systems, needs a place where its own policy is enforced before an action is allowed to happen: where a participant's identity is checked, its capability and reputation are weighed against the action it is attempting, the action is permitted or denied according to the organization's rules, and the whole interaction is recorded for whoever will later have to answer for it.

This is the control point of the new economy. It is more than an access check at the door. It is the layer where policy binds to action across every interaction a participant has, where the rules of a jurisdiction, an industry, and an individual organization all have to be evaluated and enforced together, and where the evidence that they were enforced is produced and kept. Its full set of functions is broader than any single one of these descriptions suggests, and the categories of function it has to cover include at least these: deciding whether a given action is permitted before it occurs, enforcing the limits on what a participant is authorized to commit, carrying policy across chains of delegated action so that authority cannot quietly expand as work is passed along, and producing a durable record of every decision for audit and accountability. Early self-attesting versions exist, where a platform enforces its own policy on agents running inside it. The gap is a control point that is independent of any single platform, that evaluates the policy of both sides of an interaction rather than one, and whose evidence can be trusted by a regulator, an auditor, or a counterparty who does not take the platform's word for it. The organizations that need this most are exactly the ones Section 3 identified as most exposed: regulated industries, public-sector bodies, and any enterprise whose board is now accountable for what autonomous systems do in its name. The demand is being created right now by the same regulatory codification Section 1 documented, because every obligation that lands on an organization has to be enforced somewhere.

4.7 Vertical marketplace operators

Horizontal infrastructure gets built once and serves everyone. But much of the agentic economy will organize around verticals, because the rules, the certifications, the counterparties, and the trusted practices differ sharply from one industry to the next. A marketplace where healthcare agents find, evaluate, and transact with one another operates under constraints a logistics marketplace does not share, and the reverse holds too.

The rebuild requires operators who run these vertical marketplaces: the trusted venues where participants in a given industry discover each other, verify each other's identity and standing, and transact under rules appropriate to that industry. This is a familiar pattern wearing new clothes. Industry-specific exchanges, clearing venues, and trading networks have always formed wherever a sector's transactions were specialized enough to need their own trusted intermediary. The agentic version of that intermediary is greenfield.

4.8 Specialist-agent businesses

Above the coordination categories sit the businesses that build and operate the agents themselves. Not platforms, not infrastructure, but operators whose product is a specialist agent that does a specific kind of work better than a general one, offered either as a service or for direct procurement and deployment inside the buyer's own systems for organizations whose security or regulatory posture requires the agent to run on infrastructure they control. A specialist agent for a narrow class of legal review, for a particular kind of financial reconciliation, for a specific diagnostic-support task, for a defined logistics-optimization problem.

This is the category that most resembles the application-software businesses of the previous era, and it is likely to be the most populous. The barrier to building one is low relative to the infrastructure categories beneath it, which is precisely why these businesses depend so heavily on those categories existing. A specialist-agent business is only viable if its agent can hold an identity, carry a certification, build a reputation, settle transactions, negotiate terms, and operate through a venue where counterparties will trust it. The application layer is where the most builders will go, and it is the layer that most needs everything underneath it to be in place first.

4.9 AI-optimized commerce infrastructure

Section 3 observed that the entire current retail experience, catalogs, images, reviews, search, checkout, is optimized for human cognition operating through a screen. The same is true of most commerce infrastructure across the economy. The product information, the contract terms, the pricing logic, and the fulfillment interfaces are all shaped for human readers and human decision-makers.

As agents take over the transactions where the human steps out of the loop, that infrastructure has to be rebuilt for agent consumption: structured where it was narrative, machine-checkable where it relied on human interpretation, and exposed through interfaces an agent can act on directly. The rebuild requires operators who build this agent-facing commerce infrastructure. The opportunity is large because it spans every sector where agents will transact, and almost none of it exists in agent-ready form today.

4.10 Data-as-a-service with verified provenance

Agents act on data, and the quality of an agent's action is bounded by the quality and trustworthiness of the data it acts on. In a world where data can be synthetic, manipulated, or simply wrong, and where the agent consuming it cannot independently judge its origin, provenance becomes the scarce and valuable property. A data feed an agent can trust is one whose origin, chain of custody, and integrity are provable.

The rebuild requires operators who provide data as a service with verified provenance attached: feeds whose source is identifiable, whose handling is recorded, and whose integrity an agent can check before relying on it. This is a category that grows directly out of the identity and reputation categories, because provenance is, in the end, identity and history applied to data rather than to participants. The operators who can attach trustworthy provenance to data at scale build a business on exactly the property the agentic economy is shortest on.

4.11 Household agent services

The categories above are mostly enterprise, institutional, and inter-business. But the agentic transition reaches personal life as well, and the household is its own greenfield. Individuals and families will run agents that manage household logistics, finances, health coordination, education support, and the long tail of administrative work that consumes personal time. The household needs the same things the enterprise needs, identity for its agents, certification it can trust, settlement it can rely on, the ability to negotiate on its behalf, and governance over what its agents are permitted to do, scaled down and made simple enough for a non-technical person to operate.

The rebuild requires operators who deliver agent capability as a service to the household: the consumer-facing layer of the new economy. This category barely exists today, and it is the one that will eventually touch the largest number of people, because every household is a potential participant. The operators who make agentic capability safe and simple for ordinary people to use are building toward the broadest market in the entire transition.

4.12 The list is open

Eleven categories, and the list does not close here. Every infrastructure transition produces operator categories its early observers did not anticipate, because the substrate, once it exists, enables businesses that could not have been imagined before it. The web produced search, social platforms, streaming, and the gig economy, none of them foreseeable from the vantage point of the substrate's first years. The agentic substrate will do the same.

What the eleven categories share is the thing worth holding onto. Every one of them depends on a foundation that does not yet exist in shared, interoperable form: a way for every kind of participant to be identified, to prove what it can do, to carry its history, to settle what it owes, to agree on terms, to be governed by the rules that apply to it, and to do all of this across the boundaries between one operator's system and another's. The builders in these categories are not waiting for that foundation as a convenience. They are blocked on it. The category that builds the foundation itself is the subject the rest of this document turns to.

But before it does, there is a participant in this economy that the builders cannot serve on commercial terms, because it does not operate on commercial terms at all. That is the subject of Section 5.

Section 5 — The Sovereignty Dimension

Section 4 walked the categories that get built on commercial terms. There is one participant in this economy whose involvement is not optional and is not commercial. Every jurisdiction has functions it cannot delegate, and the rebuild surfaces a set of them that have to be accommodated from the start.

Start with what a government actually owes the people it serves. Public services, infrastructure, courts, the rule of law itself, all of it rests on a state's ability to see the economic activity happening on its territory and to ask everyone who benefits from that activity to contribute a fair share to what the public depends on. That is the case for taxation, and it has nothing to do with anyone's appetite for paying tax. It has to do with whether the services and protections everyone relies on continue to exist. When two humans transact, the activity sits inside a body of tax law built up over generations, and the records the state needs to enforce it are produced as a byproduct of how humans and institutions settle with each other. When two agents transact, the same body of law applies, but the records the state needs are not produced by default, and the volume of the activity is beyond what any existing oversight mechanism was built to follow. A government in that position has not lost its authority. It has lost the visibility that makes its authority operable, and with it the ability to ensure that everyone participating in the economy contributes fairly to the public goods it makes possible.

Visibility is also what regulation depends on, where a jurisdiction has chosen to regulate. Different societies set that dial in different places. Some take a light hand and let markets operate with minimal intervention; others take a heavier one in particular sectors where the public interest is judged to warrant it. The point is not which choice a society makes. The point is that whatever choice it has made, the rules cannot be enforced against activity the state cannot see. Consumer protections, professional standards, sector-specific oversight: all of it depends on the authority of the law reaching the parties it applies to, and that authority has to land in the agentic layer in a form that fits how the activity actually happens, fast, automated, and at a volume only machines can read.

Some activities go further than regulation in the ordinary sense, because the state itself is the operator or because the public interest is judged too acute to leave to general oversight. Agentic systems are already being deployed inside defense and intelligence functions, on classified networks, for work that no jurisdiction would let a commercial party run on its own authority ⁵²⁵³. The same is true of activities a government tightly controls because of what could go wrong if it did not: research touching biological agents, gene synthesis, and other dual-use science; the development, approval, and prescription of pharmaceuticals; the handling of classified information and the systems that move it; constituent services where the state's competence and its respect for the rights of the people it serves are themselves on the line. Each of these sits inside a body of law and a body of practice that took decades to build, and each has been administered until now by humans operating under direct accountability to public authority. When agentic systems are introduced into any of them, the authority does not loosen. It tightens. And the substrate has to make that tightening enforceable, in a form the operating agency itself can rely on, not in a form a commercial vendor offers as a feature. No currently existing solution provides this. Each vendor builds what its own product needs for its own customers, on its own authority, and nothing in the open market gives a government the structural means to govern this category of activity in the way the category requires.

Cross-border movement of value is the second case, and it sharpens the first. Every jurisdiction enforces a set of rules about value moving across its borders: treaty obligations on tax cooperation, financial-crime prevention, the suppression of money laundering and criminal flows, and, where states find it necessary, even sanctions against named parties. These rules are enforced today through banking-system chokepoints, reporting requirements, and the cooperation of the institutions running the cross-border rails. Agent-to-agent value movement does not pass through those chokepoints the same way, and the rules they existed to enforce do not stop applying because the transaction is happening at machine speed. A sovereign that cannot honor its treaties, police criminal flows, or hold the line on its own border in this layer of the economy has a gap exactly where its authority is supposed to meet the rest of the world.

The broader question of how governments themselves will eventually participate, with their own digital credentials, constituent-facing services, and the modernization of the documents and permits citizens carry, is one most jurisdictions will work out over a longer horizon. Those changes are likely to come because they make sense to the governments that adopt them, but they are not the prerequisite for the rebuild. The prerequisite is the structural one this section has been describing. None of it depends on a view of how a state should be organized or where the line between public and private ought to sit. Different societies will answer those questions differently, as they always have. What is constant is that no government can let a layer of its economy sit outside the reach of the law and the tax code that govern every other layer, and the substrate has to make sure it does not.

There is precedent for problems of this shape being solved. The naming and addressing system that lets every device on the internet be found was, for years, administered under the authority of a single national government before it transitioned to a multistakeholder body with international participation. The transition was contested and imperfect, and it happened because the function mattered too much to too many parties to leave where it started. The agentic substrate raises the same kind of question at a layer that sits closer to the core of what jurisdictions actually do, which is to fund the public goods their people depend on, to set and enforce the rules they live under, and to control what crosses their borders.

The substrate the document turns to next is the one that carries these sovereign requirements as structural properties of the way it is built, rather than as obligations chased after the fact. That is the subject of Section 6.

Section 6 — Introducing Scarp

The previous two sections walked the same gap from two sides. Section 4 named the categories the rebuild creates and showed where each one stalls without something underneath it the builders can rely on. Section 5 named the functions a jurisdiction cannot delegate and showed how each one runs into the same gap from the public side. What stalls them is the absence of a substrate.

What this document proposes is a framework built on four foundational layers that together form the substrate the agentic economy needs. We call it Scarp. The four layers are Identity, Cooperation, Governance, and Settlement. Only one of them, Identity, has a recognizable starting point in the existing internet substrate, and even that one is extended substantially. The other three are built without precedent because the functions they have to perform have no precedent at this scale and in this form. That is the work Scarp is doing and what the rest of this section explains.

What a substrate is

Substrates are not new. The internet that runs every business on earth right now is one. It did not exist as a single product anyone could buy. It existed because a generation of engineers, governments, standards bodies, and operating companies built a set of agreed layers — addressing, routing, naming, transport, application protocols, trust — that everyone could rely on without any one party owning them or any one party being able to take them away. TCP/IP defined how packets move. DNS defined how names resolve to addresses. BGP defined how networks announce reachability to one another. Certificate authorities defined how a name on a certificate could be trusted to belong to the entity presenting it. Each was built because something underneath it had to exist before anything above it could function. None was built by a single company as a product, and none works because any single company runs it. They work because they are agreed, and because they were designed as structural properties of the layer they belong to rather than wrapped around it after the fact.

That is what a substrate is. It is the agreed layer underneath an economy that lets the activity above it happen at all.

The four foundational layers

Scarp proposes four foundational layers. Each one closes a gap surfaced in Sections 2 through 5. One has a recognizable starting point in the internet substrate that came before. The other three do not.

Identity

Every participant in the agentic economy — humans, organizations at every scale, governments at every layer, physical machines that operate autonomously or semi-autonomously, and the software agents acting on behalf of all of them — has to be identifiable, verifiable, and accountable. The certificate-authority framework that secures the existing internet is the starting point. The agentic regime asks substantially more of an identity layer than the web ever did. The set of participants is broader. The rate at which identities have to be issued is higher by orders of magnitude. The consequences of a compromised identity are larger, because what gets compromised is no longer a session or a credential but an autonomous actor capable of moving value, signing commitments, and taking action on someone's behalf at machine speed. And the relationship between an identity and the rules under which it operates has to be carried by the identity itself, rather than applied as external policy at runtime, because external policy is too slow and too easy to bypass at the volume the substrate has to handle. Scarp extends the certificate framework to meet those requirements. The substance of how it does so is the subject of separate technical material. What matters here is that Identity is the one layer of the substrate that already has a recognizable starting point, and that the extensions required are substantial.

Cooperation

Two agents that have no way to find each other, declare what they can do, agree on terms, and execute joint work cannot transact, coordinate, or compose into the workflows the rebuild depends on. Cooperation is the layer that lets them. It covers discovery across organizations and jurisdictions, capability advertisement under controlled disclosure, intent declaration, governed negotiation, contract formation that requires both parties' authority to be satisfied independently before any joint commitment exists, execution of joint work under bilateral governance evaluation at each step, and the structured handling of disputes when joint work does not complete as agreed. None of this has a counterpart in the internet substrate that came before. The existing internet solved how machines find each other and exchange traffic. It did not solve how autonomous actors under principal-bound authority discover each other, advertise what they can do without exposing what they are not authorized to do, negotiate to a binding agreement, and then jointly execute that agreement across organizational lines with both sides' rules enforced bilaterally. Early entrants in adjacent territory have surfaced fragments of this, and useful work is being done on agent-to-agent protocols, but nothing in the field today covers the scope. Cooperation is built fresh.

Governance

Governance is the layer that makes the activity in the substrate accountable. An AI agent operating in this economy acts on behalf of a principal who is on the hook for what it does. The operators deploying agents — banks, hospitals, manufacturers, government departments, every kind of enterprise — are on the hook for what their agents do under the laws and the contracts they operate under. Auditors, regulators, courts, counterparties, insurers, dispute-resolution bodies, and the operators and principals themselves all need a way to know what an agent actually did, against what rules, with what authority, and with what outcome — in a form none of those parties has to take on the operator's word. They also need to know whether the agent operates in a way that makes following the rules a structural property of how it acts at all, rather than a probability that a model will get the right answer at runtime. Governance in the substrate is the layer that produces that answer. It is built so that policy is bound to identity, evaluated outside the agent's reach, and recorded in a form that any qualifying party can verify independently of the operator. Self-attestation has never been accepted as the standard for regulated record-keeping in any other layer of the economy, and the courts will not accept it here either.

Settlement

The volume Section 4 surfaced — agent-to-agent value movement at orders of magnitude beyond what the existing payment and clearing infrastructure carries today — has to settle with cryptographic provenance for every transaction so that nothing reaches finality without a verifiable record of who decided, on whose authority, under what terms. Settlement in Scarp is identity-anchored at both ends, so a transaction without valid identity on both sides cannot form. Funds committed under an agreement are held under cryptographic guarantees that release them only on verified delivery or on the agreed alternative paths, with no third party able to redirect them in between. Tax remittance is a structural property of the transaction rather than a reporting obligation chased after the fact. Validators are operated by the same parties who issue identity, so the same trust hierarchy that governs who can participate also governs who can finalize. Partial functionality exists in the field today: contract-based transactions and escrow patterns have been carried on public blockchains for years, and recent work on agent-identity and agent-payment registries has begun to address the agentic regime specifically. These are useful early entrants, and they confirm the requirement is real. None of them covers the layer at substrate scope.

The four layers operate together. Identity without governance is uncertified standing. Governance without identity has nothing to bind to. Cooperation without settlement leaves work uncompensated and unenforceable. Settlement without governance moves value outside the reach of the laws every other layer of the economy operates inside. The four layers are Scarp, and the substrate is what they form when they operate together.

Why now, and why the cost of getting this wrong is higher than the internet's was

The internet protocol was not designed with security in mind. Authentication, confidentiality, and trust were not properties of the layer. They were added later, by other parties, on top. The firewall industry exists because of that omission. So does the entire enterprise security stack — intrusion detection, web application firewalls, secure web gateways, zero-trust overlays, identity-aware proxies, and the global market that now spends well over two hundred billion dollars a year retrofitting security onto a substrate that was never designed to carry it ⁵⁴. Decades of engineering, regulation, breach response, and litigation have gone into chasing fixes for a foundational gap that compounds with every new layer built on top of it.

The agentic substrate is the next layer of the world economy, not a new application running on the existing one. It will carry identity, cooperation, governance, and settlement for autonomous actors operating at machine speed across every jurisdiction and every sector simultaneously. The activity Section 1 documented is already happening. Frontier model providers, payment networks, governments, manufacturers, healthcare systems, and the standards bodies that have started to meet on agentic protocols are all running on assumptions about this substrate that do not yet hold in reality. Each is filling the gap in the way its own product needs, on its own authority. That is normal at the start of every substrate transition. It is also the period during which a substrate either gets agreed, or fragments into incompatible regional and vertical stacks with the foundational properties missing. The next generation of engineers, regulators, and operators then spends decades chasing fixes for what should have been structural from the start, instead of building applications and workflows on a solid foundation.

The cost of getting this wrong is higher than the cost of getting the internet wrong was, for a structural reason. The internet without security still ran, and security got bolted on at substantial but recoverable cost. The agentic substrate without identity, governed cooperation, accountable governance, and identity-anchored settlement does not just run insecurely. It runs without anyone being accountable for what an agent does, without verifiable evidence of what happened, without a way to enforce the laws that every other layer of the economy operates under, and without a way for any counterparty to know whether the agent on the other end of an interaction is real, authorized, or trustworthy. The economy that builds on that foundation will spend the next several decades retrofitting properties that should have been structural from the start, at scale and stakes the early internet never had to carry.

Why Scarp is the substrate

Scarp is the framework architected so the four foundational properties are structural from the start. Identity is bound at every participant. Cooperation is built fresh because no prior substrate had to carry it. Governance is the layer that makes every action accountable to the parties whose authority or interest is in scope. Settlement is identity-anchored, governed before finality, and structurally aligned with the public-interest obligations every jurisdiction enforces on every other layer of its economy. As of this writing, it is the only proposal we know of that addresses all four foundational layers as structural properties of a single substrate rather than as parts to be assembled later. That is the substrate the rebuild needs, and that is what Scarp is.

What this section has done, and what comes next

This section has named the substrate, placed its four foundational layers, and made the case that the cost of leaving the load unaddressed compounds rapidly across the economy that will be built on top of it. It has not specified how Scarp will be built, owned, governed, licensed, or operated. Those questions are real, and they are open. They are answered the way substrate questions have always been answered: through the bodies, conventions, and agreements that bring together the parties whose decisions actually determine the shape. The next section names the path Scarp is taking and what is open about how it gets built.

Section 7 — The Standards Pathway

The previous section closed by naming that the questions of how Scarp gets built, owned, governed, licensed, and operated are real, they are open, and they are answered the way substrate questions have always been answered — through the bodies, conventions, and agreements that bring together the parties whose decisions actually determine the shape. This section is where those questions get walked.

The choice

Scarp is choosing the standards-pathway route. That is the choice this document is putting on the table. The work develops as a standard, in an open multistakeholder venue, with the parties whose decisions determine deployment present at the table, with the technical material refined under open review by those parties, and with the governance of the standard living outside any single party once the venue is constituted. That is the route the substrate is taking. It is not chosen by elimination of alternatives and it is not chosen because no other path was available. It is chosen because that is the route a substrate of this scope and this consequence has to be built on if the parties who will rely on it are going to be able to rely on it.

What is open

The shape of how the standard gets built and governed is a separate set of questions, and those are open. The venue could be one of the existing standards bodies extending its scope. It could be a new body constituted specifically for the agentic substrate. It could be a federation of bodies, each carrying the layers most adjacent to the work each has historically done. The composition of the parties at the table — which incumbents, which builders, which governments, which civil-society participants — is open. The foundation form, if a foundation is the right vehicle, is open. The governance composition is open. None of those are foreclosed by anything the document has said. They are the shape questions, and the shape questions are answered by the parties who join the work.

The four foundational layers — Identity, Cooperation, Governance, Settlement — are what the standard has to carry, in the form the previous section described. The work of refining how they are specified, how they interoperate, how they evolve, and how they bind to the parties who implement them is the work the venue is for. The technical material that develops the layers in detail exists alongside this document, held under appropriate confidentiality until parties engaging with the substrate are in a position to receive it under terms suited to that engagement.

The call to participate

The call is open to the parties whose decisions determine the substrate's shape. To incumbents whose existing infrastructure will interconnect with the substrate, to builders whose work will rest on it, to governments whose regulatory and sovereign frameworks will need to reference and rely on it, to standards-body participants whose experience carrying prior substrates is the experience this one will need, and to investors and convenors whose capacity to bring parties to the table is part of how the venue actually forms.

The intellectual-property regime under which the substrate gets licensed and implemented is open. Patent applications covering the substrate's foundational claim surface are on file and pending, comprising more than one hundred claims. The disposition of those claims — how they are licensed, donated, retained, or otherwise structured for the standard's adoption — is one of the questions the venue conversation will work through, and the document is not the place to commit to a regime before that conversation has happened.

Section 8 — What Comes Next

The document leaves the reader at the threshold of a conversation, not at the end of an argument. The substrate has been named. The four foundational layers have been described. The path Scarp is taking has been declared, and the questions of shape that the path will work through have been listed openly. What remains is the conversation itself, and the parties who carry it forward.

What that conversation looks like, from this side of the document, is straightforward. Serious engagement starts with a direct exchange — a message, a call, a meeting. Parties who recognize the substrate's significance, who see their own infrastructure or mandate or work intersecting with what is being proposed, and who are positioned to contribute to how the substrate takes shape are the parties this document is written for. The conversation that follows is the conversation that determines the venue, the governance, the parties at the table, and the structural arrangements through which the substrate gets built. Those questions are not abstract. They get answered through specific exchanges with specific parties, and those exchanges begin with the first message.

What that conversation does not look like is also worth naming. It is not a sales process, and the document is not soliciting customers. It is not a fundraise in the conventional sense, though investors and convenors with the capacity to assemble the venue are part of who the document is addressed to. Parties who engage are engaging at a stage where the structural arrangements are still being formed, and the form they take will reflect the parties who shape them.

For those parties, the contact is direct.

victor@scarpprotocol.com

References

[1] METR. "Task-Completion Time Horizons of Frontier AI Models." Updated periodically. https://metr.org/time-horizons/ ↩

[2] METR. "Time Horizon 1.1." Published January 29, 2026. https://metr.org/blog/2026-1-29-time-horizon-1-1/ ↩

[3] OpenAI. "Introducing SWE-bench Verified." Published August 13, 2024. https://openai.com/index/introducing-swe-bench-verified/ ↩

[4] SWE-bench. "SWE-bench Verified Leaderboard." Updated periodically. https://www.swebench.com/verified.html ↩

[5] Scale AI. "SWE-Bench Pro Leaderboard." Updated periodically. https://labs.scale.com/leaderboard/swe_bench_pro_public ↩

[6] KPMG. "AI Quarterly Pulse Survey." Series, Q2 / Q3 / Q4 2025. https://kpmg.com/us/en/articles/2025/ai-quarterly-pulse-survey.html ↩1 ↩2 ↩3

[7] CrewAI. "2026 State of Agentic AI Survey Report." Published February 2026. https://www.businesswire.com/news/home/20260211693427/en/ ↩1 ↩2

[8] Cleanlab. "AI Agents in Production 2025: Enterprise Trends and Best Practices." Survey published 2025. https://cleanlab.ai/ai-agents-in-production-2025/ ↩

[9] InsuranceNewsNet, 2025 industry analysis, as reported in Multimodal, "10 AI Agent Statistics for 2026," December 22, 2025. https://www.multimodal.dev/post/agentic-ai-statistics ↩1 ↩2

[10] AgentMarketCap. "AI Captured 61% of Global Venture Capital in 2025 — And the Concentration Is Getting Worse." Published April 10, 2026. https://agentmarketcap.ai/blog/2026/04/10/ai-captures-50-percent-global-venture-capital-2025 ↩1 ↩2 ↩3 ↩4 ↩5 ↩6

[11] TechCrunch. "Here are the U.S. AI startups that raised $100M or more in 2025." Published January 19–20, 2026. https://techcrunch.com/2026/01/19/here-are-the-49-us-ai-startups-that-have-raised-100m-or-more-in-2025/ ↩1 ↩2 ↩3 ↩4 ↩5

[12] Menlo Ventures. "2025: The State of Generative AI in the Enterprise." Published March 19, 2026. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/ ↩1 ↩2

[13] KPMG. "AI Quarterly Pulse Survey, Q1 2026 release including AI spending projections." Same series as [6]. ↩1 ↩2

[14] CB Insights. "State of AI 2025 Report." Published February 5, 2026. https://www.cbinsights.com/research/report/ai-trends-2025/ ↩1 ↩2 ↩3 ↩4

[15] Anthropic. "Model Context Protocol release announcement." November 25, 2024. As described in Wikipedia, "Model Context Protocol." https://en.wikipedia.org/wiki/Model_Context_Protocol ↩

[16] ByteBridge. "Model Context Protocol (MCP): Evolution, Capabilities, and the Rise of Peta." Published January 4, 2026. https://bytebridge.medium.com/model-context-protocol-mcp-evolution-capabilities-and-the-rise-of-peta-ff2967b45d48 ↩

[17] Imagine Works. "Model Context Protocol (MCP): A Guide for Enterprise Leaders." Published April 20, 2026. https://imagine-works.com/insights/model-context-protocol-for-enterprise-leaders ↩1 ↩2

[18] Neosalpha. "Model Context Protocol Explained: What Is MCP and Why It Matters." Published May 2026. https://neosalpha.com/blogs/what-is-mcp/ ↩

[19] Digital Applied. "MCP Hits 97M Downloads: Model Context Protocol Guide." Published March 9, 2026. https://www.digitalapplied.com/blog/mcp-97-million-downloads-model-context-protocol-mainstream ↩

[20] Wikipedia. "Agent2Agent." https://en.wikipedia.org/wiki/Agent2Agent ↩1 ↩2

[21] Wikipedia. "Model Context Protocol," section on Linux Foundation donation. https://en.wikipedia.org/wiki/Model_Context_Protocol ↩1 ↩2

[22] Wikipedia. "Agentic commerce," section on Universal Commerce Protocol. https://en.wikipedia.org/wiki/Agentic_commerce ↩1 ↩2 ↩3 ↩4 ↩5

[23] Ecma International. "Ecma International approves NLIP standards suite for universal AI agent communication." Published December 10, 2025. https://ecma-international.org/news/ecma-international-approves-nlip-standards-suite-for-universal-ai-agent-communication/ ↩1 ↩2

[24] European Union. "Regulation (EU) 2024/1689 (the EU AI Act)." Entered into force August 1, 2024. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai ↩1 ↩2

[25] Legiscope. "EU AI Act Timeline: Key Dates and Deadlines." Updated May 2026. https://www.legiscope.com/blog/eu-ai-act-timeline-deadlines.html ↩1 ↩2

[26] European Commission. "AI Office." https://digital-strategy.ec.europa.eu/en/policies/ai-office ↩

[27] The White House. "Executive Order 14179: Removing Barriers to American Leadership in Artificial Intelligence." January 23, 2025. https://www.whitehouse.gov/ ↩

[28] The White House. "Winning the Race: America's AI Action Plan." Published July 2025. https://www.whitehouse.gov/ ↩

[29] The White House. "Ensuring a National Policy Framework for Artificial Intelligence." Executive order issued December 11, 2025. https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/ ↩

[30] White & Case LLP. "AI Watch: Global Regulatory Tracker — United States." Updated periodically. https://www.whitecase.com/insight-our-thinking/ai-watch-global-regulatory-tracker-united-states ↩

[31] UK Government. "AI Regulation White Paper." Available via GOV.UK. https://www.gov.uk/ ↩

[32] Royal Society Open Science. "Artificial intelligence policy worldwide: a comparative analysis." Published February 1, 2026. https://royalsocietypublishing.org/rsos/article/13/2/242234/480264/ ↩1 ↩2

[33] Singapore Government Technology Agency. "Agentic AI Primer." Published 2025. Available via GovTech Singapore. https://www.tech.gov.sg/ ↩1 ↩2 ↩3

[34] Stanford HAI. "2026 AI Index Report." Published 2026. https://hai.stanford.edu/ai-index ↩1 ↩2 ↩3

[35] OECD AI Policy Observatory. https://oecd.ai/ ↩1 ↩2

[36] Fortune, citing FactSet. "Earnings calls citing 'AI' surge in 2025 as 'uncertainty' mentions fade." Published December 17, 2025. https://fortune.com/2025/12/15/earnings-calls-citing-ai-surge-2025-uncertainty-mentions-fade-cfo/ ↩

[37] Salesforce. "Building the Agentic Enterprise: Salesforce News and Stories That Shaped 2025." Published December 23, 2025. https://www.salesforce.com/news/stories/2025-recap/ ↩1 ↩2 ↩3 ↩4 ↩5 ↩6

[38] Fortune. "CEO of a $25.9 billion bank had his AI clone lead the company's earnings call." Published April 28, 2026. https://fortune.com/2026/04/28/ceo-of-customers-bank-sam-sidhu-ai-clone-lead-earnings-call-mark-zuckerberg-building-digital-twin/ ↩1 ↩2 ↩3 ↩4 ↩5

[39] CNBC. "This bank CEO let his AI clone handle an earnings call — now he's signing an OpenAI deal." Published April 27, 2026. https://www.cnbc.com/2026/04/27/openai-partners-with-customers-bank-in-push-to-automate-finance.html ↩1 ↩2 ↩3 ↩4

[40] IoT Analytics. "What CEOs talked about in Q2 2025." Published September 25, 2025. https://iot-analytics.com/what-ceos-talked-about-in-q2-2025-tariffs-recession-agentic-ai/ ↩1 ↩2 ↩3

[41] Brave New Coin. "AI Agents Anticipated to Enter Workforce in 2025, Says OpenAI CEO." Published January 8, 2025. https://bravenewcoin.com/insights/ai-agents-anticipated-to-enter-workforce-in-2025-says-openai-ceo ↩1 ↩2

[42] SHRM. "The AI Layoffs Narrative: Real Transformation, or Scapegoat?" Published May 2026, citing Challenger, Gray and Christmas tracking. https://www.shrm.org/topics-tools/news/technology/ai-layoffs-transformation-scapegoat ↩

[43] Cisco. "Cisco Expands AgenticOps Innovations Across Portfolio." Published February 10, 2026 (Cisco Live EMEA, Amsterdam). https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m02/cisco-expands-agenticops-innovations-across-portfolio.html ↩1 ↩2

[44] Cisco. "The Agentic AI Era Demands a New Network." Cisco Blogs, October 14, 2025. https://blogs.cisco.com/networking/the-agentic-ai-era-demands-a-new-network ↩1 ↩2

[45] Cisco. "Cisco Crosswork Multi-Agentic AI Framework At-a-Glance." https://www.cisco.com/c/en/us/products/collateral/networking/software/crosswork-network-automation/crosswork-multi-agentic-ai-framework-aag.html ↩

[46] The Imaging Wire. "FDA Updates AI List with New Clearances." Published March 11, 2026, tracking authorizations through end of December 2025. https://theimagingwire.com/2026/03/11/numbers-from-the-fda-show-radiology-is-maintaining-its-lead/ ↩1 ↩2

[47] Sivakumar R et al., "AI in radiology and interventions: a structured narrative review of workflow automation, accuracy, and efficiency gains." PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC12929352/ ; supporting market projection from IntuitionLabs, "AI in Radiology: 2025 Trends, FDA Approvals & Adoption." https://intuitionlabs.ai/articles/ai-radiology-trends-2025 ↩1 ↩2

[48] Visa Inc. Fiscal 2024 Annual Report (Form ARS), filed with the U.S. Securities and Exchange Commission. Fiscal 2024: approximately 233.9 billion transactions processed on VisaNet; 303 billion total Visa-brand payments and cash transactions; average of 829 million transactions per day. https://www.sec.gov/Archives/edgar/data/0001403161/000130817924000809/v4356631-ars.pdf ↩

[49] Swift. Annual Review 2024. 13.4 billion FIN messages in 2024; daily average of 53.3 million; record daily FIN traffic of 59.5 million (20 December 2024). https://www.swift.com/sites/default/files/files/swift-annual-review-2024.pdf ↩

[50] IEEE Spectrum. "AI Agents Are Taking Over Contract Negotiations." Reporting enterprise deployment of autonomous supplier-contract negotiation systems at large global enterprises. https://spectrum.ieee.org/ai-contracts ↩

[51] Stanford HAI / Stanford Digital Economy Lab. "The Art of the Automated Negotiation," reporting on the study "The Automated but Risky Game: Modeling Agent-to-Agent Negotiations and Transactions in Consumer Markets." Finds wide capability disparities between negotiating agents and inconsistent adherence to user-set constraints. https://hai.stanford.edu/news/the-art-of-the-automated-negotiation ↩

[52] U.S. Department of War. "Classified Networks AI Agreements." Press release, May 2026. Announces agreements with eight frontier AI firms to deploy AI capabilities on the Department's Impact Level 6 (information classified up to Secret) and Impact Level 7 (highly restricted) classified networks for lawful operational use. https://www.war.gov/News/Releases/Release/Article/4475177/classified-networks-ai-agreements/ ↩

[53] U.S. Department of Commerce, Bureau of Industry and Security. Controls on Certain Laboratory Equipment and Related Technology To Address Dual Use Concerns About Biotechnology, Interim Final Rule, effective January 16, 2025. Federal Register. Imposes export controls on biotechnology equipment, gene-synthesis tools, and related technology under dual-use concerns including the training of AI systems for militarily relevant applications. Companion AI Diffusion Rule (ECCN 4E091) controls model weights of advanced closed-weight dual-use AI models. https://www.federalregister.gov/documents/2025/01/16/2025-00723/controls-on-certain-laboratory-equipment-and-related-technology-to-address-dual-use-concerns-about ↩

[54] Gartner. Forecast: Information Security and Risk Management, Worldwide, 2022–2028, 2Q24 Update (and successor quarterly updates). Global end-user spending on security and risk management projected at approximately $212 billion in 2025, up from approximately $193.5 billion in 2024. Used here as the conservative anchor for the global enterprise-security market figure. https://www.gartner.com/en/newsroom/press-releases/2024-08-28-gartner-forecasts-global-information-security-spending-to-grow-15-percent-in-2025 ↩