State of the agent economy

Agenstry’s observed agent web expanded materially in ISO week 2026-W37: 8,041 candidates were indexed, up +559.00 (+7.5%) week on week; the live, introspectable layer reached 734, up +140.00 (+23.6%); and 2,486 agents showed inbound stablecoin settlement over the measured 30-day window, up +286.00 (+13.0%). The central qualification is commercial depth: observed 30-day gross settlement was 244999.43 USD, but revenue remained highly concentrated, with the top agent accounting for 72.9% and the top ten for 95.6%.

Market structure: more endpoints, still a thin verified runtime layer

The funnel continues to show the difference between discoverability and operational readiness. Of 8,041 indexed candidates, 6,475 returned a parseable card, while 734 were live JSON-RPC or introspectable MCP endpoints. Only 98 published a JWS-signed card. That makes signed identity the narrowest measured layer of the agent web, even as endpoint discovery accelerates.

The paid surface is broader than runtime verification. Agenstry labels 4,397 payment-capable records, but only 335 are self-declared live through an endpoint that returned an unpaid HTTP 402 challenge or served a /.well-known/x402 manifest; the stated measurement defines runtime payment challenges as authoritative. The implication for buyers, registries and regulators is straightforward: catalog presence should remain distinct from a tested ability to quote, authenticate and settle a transaction.

x402 remains the default settlement vocabulary in the labelled corpus, with 4,243 records, versus 88 for MPP, 40 for A2A+x402 and 21 for AP2. The ISO week 2026-W37 cohort is necessarily incomplete as of the 2026-09-07T13:30:42.229154+00:00 measurement time: it contains 37 additions, all x402-labelled, and should not be read as a full-week deceleration from the 406 additions recorded for 2026-W36.

Where the money is flowing

The revenue distribution is more concentrated than the growth figures alone imply. Across 2,486 earning agents and 1,256 earning wallets, observed 30-day gross settlement was 244999.43 USD; the median was 0.26 USD, while the p95 was 44.6397 USD. The Gini coefficient was 0.994, and the HHI rose from 4,934.10 to 5,487.40, a +553.30 (+11.2%) week-on-week increase.

That pattern is consistent with an early market in which a small number of endpoints capture most recorded settlement while a long tail tests pricing, distribution and machine-payable interfaces. The highest-priority demand gap remains audio transcription: Agenstry records 231 variants against 663 unique sources, a 50% unmet share and a 2264.7% rising trend. This is an actionable gap because it is a bounded, readily benchmarked service with a clear per-call pricing model rather than a general-purpose agent category.

Spec and protocol: production concerns are moving from connectivity to operability

MCP’s most consequential recent change remains the final 2026-07-28 specification: it removes protocol-level sessions and initialization, uses self-describing requests, adds header-based routing and cacheable list responses, formalizes extensions, and hardens authorization. The maintainers’ subsequent roadmap prioritizes agentic messaging, HTTP-native transport, identity and enterprise security, suggesting that implementation consistency—not a new core transport—is now the central standards problem. MCP’s July specification announcement and its August roadmap support that reading. (github.com)

On payments, the x402 codebase recently introduced a settlement_pending outcome for transactions broadcast but not yet confirmed, allowing clients to reconcile an uncertain transaction rather than treating it as a terminal failure. The official changelog also records security fixes to payment-gate route matching. (github.com) A reported issue opened on September 1 alleges that Coinbase’s hosted facilitator rejected previously settling Solana exact-payment envelopes following a configuration or deployment change; it remains an open issue, not a confirmed protocol-wide defect. The issue record is nevertheless a useful reminder that live facilitator behaviour must be separately probed from SDK conformance. (github.com)

Notable deployments and research

Enterprise launches this week favoured governed, domain-specific workflow agents rather than open agent marketplaces. CBTS launched Forge Agents for mid-market deployment and management on September 1, while ProcessUnity introduced agents for third-party risk-management work that retains human judgment for escalation. CBTS’s launch announcement and ProcessUnity’s release underline the commercial premium placed on security, operational control and bounded authority. (cbts.com)

Research is converging on the same implementation constraint. A September 4 review argues that interoperability layers such as MCP and A2A improve connection between systems but do not themselves establish trustworthy delegation, emphasizing provenance, bounded authority, failure detection and recovery. “From Language Models to World-Acting Systems” is a useful synthesis for purchasers assessing whether an “agent-ready” interface actually supports accountable action. (arxiv.org)

Regulation

The regulatory direction is toward auditable control of delegated actions. The EU AI Act’s general application date was 2 August 2026, although subsequent EU legislation delays specified high-risk-system obligations; firms selling agents into Europe should distinguish those timelines rather than assume a single compliance date. EUR-Lex’s AI Act summary and Regulation (EU) 2026/1744 set out the split. (eur-lex.europa.eu) In the United States, reporting this week identified a proposed Stop Rogue AI Act focused on NIST standards and deployment practices for AI agents; it should be treated as an early legislative signal rather than operative law. Axios’s report provides the current public account. (axios.com)

What to watch next week

The key test is whether the gain in live endpoints converts into more authoritative payment evidence: specifically, whether the count of agents producing observed HTTP 402 challenges rises alongside the broader index. Investors and operators should also monitor resolution of the hosted-facilitator issue, because settlement reliability—not discovery volume—is the near-term constraint on recurring agent-to-agent commerce.