State of the agent economy

The agent economy is adding supply faster than it is proving operational reliability or broad-based monetisation. In ISO week 2026-W40, Agenstry indexed 10,167 candidates, up +514.00 (+5.3%) from 2026-W39, while the live population rose +97.00 (+11.5%) to 943 and earning wallets rose +224.00 (+17.2%) to 1,529; revenue remains highly concentrated despite a lower HHI.

Market quality is improving, but verification remains the constraint

Of 10,167 indexed candidates, 7,753 returned a valid card, 943 were live, and 129 were signed. The distinction matters: directory presence is not execution readiness, and a declared payment interface is not equivalent to a runtime payment challenge. Of the 5,328 payment-labelled services, only 371 were self-declared live through an unpaid HTTP 402 response or a /.well-known/x402 manifest; that is 7.0 of labelled services by the supplied measurement definition.

The largest source pools are not necessarily the most reliable. a2aregistry recorded 77.1 live from 301 seen, while registry recorded 58.2 live from 598 seen; by contrast, the larger agentic_market pool recorded 2.7 live from 3,936 seen. This is an operational argument for continuous probing rather than catalogue-led discovery. The W40 protocol cohort currently shows 27 additions, all x402, versus 324 in W39; however, W40 began on 2026-09-28 and the dataset is timestamped 2026-09-28T13:29:43.210629+00:00, so it is a partial-week observation rather than a meaningful deceleration signal.

Where the money is flowing

Observed gross settlement was 814624.22 USD over 30 days across 3,012 earning agents, 1,529 earning wallets, and 1,188 active sellers. The median was 0.44 USD and the p95 was 102.0 USD, while the top one, five and ten sellers accounted for 75.0, 91.4, and 95.5 of gross volume respectively. The Gini was 0.995 and HHI was 5703.5.

The week-over-week picture is mixed but constructive. The HHI declined by -1,115.50 (-16.4%) from 6,819.00 to 5,703.50, indicating that the observed settlement base became less concentrated by that measure, even as the supplied Gini comparison remained 0.99 to 0.99. That is not evidence of a competitive long tail yet; it is evidence that a small number of non-leading sellers captured more observed flow. The relevant next test is whether this persists while the 701 new earning agents recorded over 30 days mature into repeat sellers.

Spec & protocol: implementation is moving from declaration to conformance

x402 remains the dominant payment declaration: 5,160 services are recorded under x402, against 90 under MPP, 48 under A2A-x402, 25 under AP2, 4 under L402, and 1 under Stripe SPT. This installed-base lead should not be conflated with proof of a live payment endpoint; runtime payment challenges remain the authoritative test.

The standards stack is nevertheless becoming more explicit about migration and verification. MCP’s published roadmap frames the July 28, 2026 specification as the main recent release, while its changelog formalises deprecations including Roots, Sampling, Logging, HTTP+SSE, and OAuth Dynamic Client Registration in favour of newer mechanisms. MCP roadmap MCP changelog (blog.modelcontextprotocol.io) The official Go SDK has implemented the 2026-07-28 line and its conformance-suite integration, a concrete sign that protocol revision is entering deployable tooling rather than remaining a governance exercise. MCP Go SDK releases (github.com)

For A2A, the latest public repository release remains 1.0.1, dated May 26, 2026, with HTTP-binding and task-status fixes; there was no newer public release surfaced in this week’s review. A2A releases (github.com) AP2’s current public specification describes signed checkout and payment mandates, including autonomous “Human Not Present” flows, making it principally an authorisation-and-evidence layer rather than a settlement rail. AP2 specification (ap2-protocol.org)

Capital and regulation are converging on control layers

This week’s capital activity favoured enterprise deployment and runtime control. Magentic announced an $18 million Series A for industrial digital workers; Ema announced a $77 million Series B for enterprise AI employees; and runtime-security company Kontext announced $4 million in financing. Magentic announcement Ema announcement Kontext announcement (magentic.com) The inference is that buyers and investors are placing growing value on governing agent actions, not only building agent interfaces.

Regulatory attention is catching up to settlement infrastructure. On September 24, the Federal Reserve requested comment on proposed capital, reserve, risk-management and application rules for Board-supervised payment-stablecoin issuers under the GENIUS Act. Federal Reserve release (federalreserve.gov) Separately, EMVCo’s draft framework for card-based agentic payments focuses on identifying agents, communicating relevant attributes, and signalling agent-mediated transactions; feedback closes September 30. EMVCo framework announcement (emvco.com)

Research is sharpening the threat model

A recent formal analysis modelled x402, MPP, ACP and AP2, arguing that security guarantees must bind delegated authority to downstream economic and service effects across participants and lifecycle stages. “A Formal Analysis of Agent Payment Protocols” (arxiv.org) This supports Agenstry’s measurement hierarchy: endpoint liveness, verifiable identity, payment challenge behaviour and observed settlement should be treated as separate properties, not collapsed into a single “agent is live” label.

What to watch next week

The specific item to track is the response to EMVCo’s September 30 feedback deadline. If card-network participants converge on interoperable agent identity, mandate and transaction-signalling requirements, the commercial bottleneck may shift from discovering payment-capable agents to proving who authorised each action and who bears liability when it fails.