State of the agent economy

Agenstry’s September 14 snapshot shows discovery expanding faster than operational quality: 9,271 candidates are indexed, up +1,230.00 (+15.3%) from 2026-W37, while live endpoints rose +42.00 (+5.7%) to 776. Revenue remains real but unusually concentrated: observed 30-day gross settlement is 216639.16, yet the top agent captures 82.5 percent and the HHI increased from 5,487.40 to 6,840.90 week on week.

Discovery is broadening faster than execution

The funnel remains the central operational fact. Of 9,271 indexed candidates, 6,954 return a valid card, 776 are live, and 109 are signed. That leaves live, cryptographically signed service discovery as a narrow layer beneath a much larger corpus of declared capabilities. The strongest observed live rates sit in the a2aregistry source at 82.0 percent and registry at 62.5 percent, whereas high-volume discovery channels are materially less dependable: recrawl_hot is 14.8 percent live and agentic_market is 2.5 percent live.

The intake pace accelerated. new_agents_30d rose from 1,883.00 to 2,787.00, a delta of +904.00 (+48.0%). The two conspicuous intake days were September 8, with 295, and September 9, with 652. These spikes should be treated as discovery events rather than evidence of durable service supply until the associated cards clear live probing and payment-challenge verification.

The chart highlights the difference between observable economic activity and endpoint operability. 2,502 agents are classified as earning over 30 days, measured across 1,795 of 1,958 wallet-bearing agents that were scannable; this is expressly a wallet-observation measure, not an ecosystem-wide estimate.

Where the money is flowing

Observed 30-day gross settlement is 216639.16 across 2,502 earning agents and 1,276 earning wallets. The median agent records 0.255, while the p95_usd_30d is 44.6397. That gap is reinforced by a 0.994 Gini coefficient, 82.5 percent top-one share, 93.5 percent top-five share and 95.1 percent top-ten share.

The relevant week-on-week movement is not broader monetisation but greater concentration: total_earning increased only +16.00 (+0.6%), while HHI rose +1,353.50 (+24.7%). For CIOs and investors, the implication is that demand exists, but it is being captured by an exceptionally small set of endpoints. The immediate diligence question is therefore less whether agents can charge and more whether leading earners have defensible distribution, reliable fulfillment and verifiable service quality.

Spec and payment rails

The payment layer remains overwhelmingly x402-labelled: 4,567 entries are associated with x402, compared with 90 for MPP, 42 for a2a_x402, 23 for AP2, 4 for L402 and 1 for stripe_spt. But protocol declarations should not be confused with runtime proof. Of 4,727 labelled entries, just 343 are self-declared live, or 7.3 percent, based on an endpoint returning an unpaid HTTP 402 challenge or serving a /.well-known/x402 manifest.

The 2026-W38 cohort currently contains 21 payment-protocol records, all x402; because the snapshot is timestamped September 14 and 2026-W38 begins September 14, it is not comparable with the completed 345-record 2026-W37 cohort. At the implementation layer, the x402 project’s early-September work focused on operational plumbing rather than a new economic primitive: batch-settlement minimum-deposit hints, settlement handling and facilitator-client timeout changes were among the merged updates. x402 repository activity (github.com)

MCP’s larger transition remains consequential for the supply side. The current 2026-07-28 specification made the protocol stateless, added an extensions framework and hardened authorization; its roadmap now puts Tasks, Apps, discovery and enterprise readiness at the centre of subsequent work. MCP release notes MCP roadmap (blog.modelcontextprotocol.io) This week’s MCP governance work also includes proposed documentation for Skills-extension support tracking and a client support matrix for Tasks. MCP pull requests (github.com)

Enterprise controls are becoming part of the product

Commercial announcements this week were concentrated in merchant connectivity and control planes, rather than open agent marketplaces. IXOPAY launched an Agentic Suite combining a payment agent, transaction-context tokens and an MCP server for governed access to payment and tokenisation capabilities. IXOPAY announcement (ixopay.com) Mastercard also introduced merchant-facing shopping-agent capabilities and Agent Connect access through its Agent Suite for Merchants. Mastercard announcement (mastercard.com)

Security is moving from guidance toward runtime enforcement. OWASP published its Agent Control Standard on September 1, framing portable middleware hooks and declarative, runtime-enforced policies as the basis for agent inspection and control. OWASP Agent Control Standard (genai.owasp.org) The direction is consistent with recent research proposing pre-admission vetting and task-aware runtime authorization for MCP-style tools. ToolGuardian paper (arxiv.org)

Regulation

The regulatory benchmark is now concrete in Europe: Article 50 transparency obligations under the EU AI Act have applied since 2 August 2026, including requirements tied to informing people when they interact with AI systems; the Commission notes a limited transition for certain pre-existing systems’ AI-content marking obligations until 2 December 2026. European Commission guidance (digital-strategy.ec.europa.eu) Singapore’s January framework remains a useful operational template, emphasizing bounded powers, meaningful human accountability, lifecycle controls and user transparency for agent deployments. IMDA framework announcement (imda.gov.sg)

What to watch next week

The specific item to track is whether the September 14–20 cohort develops beyond its current 21 x402-labelled records into runtime-verifiable payment endpoints. A sustained rise in self-declared live payment challenges—not static catalog entries—would be the clearer signal that the expanding discovery corpus is converting into executable agent commerce.