State of the agent economy

The decisive signal in 2026-W33 is not discovery volume but the gap between discoverability, runtime availability and monetisation. Agenstry indexed 6204 candidates, up +349.00 (+6.0%) week on week, while live endpoints rose to 435, up +49.00 (+12.7%); observed earning agents rose more sharply to 1467, up +407.00 (+38.4%). Yet the revenue base remains exceptionally concentrated: the top one agent accounts for 83.5% of observed 30-day gross settlement, while the Gini reached 0.994 and HHI 7032.4.

Measurement: discovery is scaling faster than verifiable operations

The corpus now contains 6204 indexed candidates. Of those, 4657 returned a parseable A2A/MCP card, 435 were live JSON-RPC or introspectable MCP endpoints, and 55 published a JWS-signed card. The practical implication is that agent discovery remains considerably ahead of deployable, authenticated runtime infrastructure: a large share of supply can describe itself, but far fewer services can be reached and interrogated as functioning agent endpoints.

The largest directly productive cohort is not identical to the live-endpoint cohort. 1467 agents were observed receiving inbound USDC/EURC settlement to declared x402 pay-to wallets over the 30-day window, measured across 1177 scanned wallets out of 1289 scannable wallets. That is a wallet-observation measure rather than an ecosystem-wide revenue census, and should not be read as evidence that every earning endpoint is currently introspectable.

Source quality is also uneven. 3227 protocol-labelled records exist, but only 233, or 7.2%, are self-declared live through an unpaid HTTP 402 challenge or a /.well-known/x402 manifest; runtime payment challenges remain the stronger operational signal. The high live rates in a2aregistry (77.2) and registry (54.4) contrast sharply with cold recrawls (0.4) and n8n workflows (0.0), suggesting that curated directories are more useful for verification than indiscriminate catalog ingestion.

Where the money is flowing

Observed gross settlement over the trailing 30 days was 123086.13 across 1467 earning agents and 695 earning wallets. The median agent recorded 0.2, versus 34.8073 at the 95th percentile. The top five agents captured 96.3% of gross settlement and the top ten 97.2%. This is not yet a broad-based services market: it is an early transactional network with a very small number of dominant settlement destinations.

Concentration has intensified since 2026-W32. The supplied week-over-week Gini delta is +0.03 (+3.1%), and HHI increased +1,791.40 (+34.2%). That pattern matters more than the headline rise in earning-agent observations. It indicates that new wallets and declared payment endpoints are appearing more rapidly than independently scaled demand is being distributed across suppliers.

Two demand gaps stand out. “Image generation” is labelled active, with 3 variants, 30 unique sources and 56% unmet demand. “Bitcoin payment” is labelled very active, with 34 variants, 93 unique sources and 49% unmet demand. Both are potentially attractive categories for new supply, but the difference between observed agent counts and concentrated settlement argues for validating paid execution—not merely publishing cards or catalog entries—before underwriting a market opportunity.

Spec & protocol

x402 remains the dominant payment declaration in the corpus, with 3103 records, compared with 84 for MPP, 19 for AP2, 17 for A2A+x402, 3 for L402 and 1 for Stripe SPT. In the 2026-W33 cohort, 18 of 21 records were x402 and 3 AP2. This should be treated cautiously: the W33 data is as of 2026-08-10T13:29:28.676666+00:00, so it represents the opening portion of the ISO week rather than a completed seven-day comparison with W32’s 225 total.

The consequential external protocol event remains MCP’s 2026-07-28 release. It replaces the protocol’s session-oriented core with stateless request/response operation, introduces Multi Round-Trip Requests for server-to-client interaction, adds header-based routing, cacheable list results, an extensions framework and authorization changes around issuer validation and client metadata. The design should lower the operational friction of horizontally scaling remote MCP servers, while requiring operators to manage a non-trivial migration away from session assumptions. MCP’s release notes and GitHub’s early support announcement indicate that ecosystem adoption is already moving ahead of the official release date. (blog.modelcontextprotocol.io)

A2A’s institutional footing continues to matter for discovery: the protocol is now housed within the Linux Foundation project structure and includes official SDKs, sample implementations, an inspector and a technology-compatibility kit. The project’s public repository frames A2A as an interoperability layer for independent agent systems, rather than a payment or governance standard. (github.com)

Regulation and enterprise controls

The regulatory clock has become more concrete. The European Commission’s AI Act Service Desk says that, from 2 August 2026, Article 50 transparency rules apply where AI agents are intended to interact with natural persons or generate content; it also states that agents are covered through the Act’s existing definitions rather than through a standalone legal category. The Commission guidance makes agent identity, disclosure and evidence trails more than voluntary design features for customer-facing deployments. (ai-act-service-desk.ec.europa.eu)

Payments standards are moving in the same direction. Google transferred AP2 to the FIDO Alliance in April, positioning the protocol within a broader standards body focused on secure authentication; AP2’s model centres on proving the authority granted to an agent for a payment action. Google’s announcement is therefore strategically relevant to the Agenstry finding that payment declarations are common but signed service identities remain scarce. (blog.google)

What to watch next week

The key test is whether the MCP 2026-07-28 migration produces a measurable rise in genuinely live, challenge-verifiable services rather than another expansion in static cards and catalog listings. Particular attention should go to whether AP2 and MPP records gain runtime evidence, and whether the increase in earning-agent observations begins to diffuse beyond the small set of wallets currently dominating settlement.