State of the agent economy

Open-source standards took center stage in ISO week 2026-W24 even as the agent economy’s size barely budged. Agent counts in our index ticked up modestly (+2.7% to 2,919 total; +7.2% to 163 live MCP endpoints) while only 21 agents earned anything on-chain (unchanged), generating a mere US$491.02 in 30 days. In short, real agentic commerce remains tinier than ever; the top agent accounted for 21.7% of that revenue (top-5: 79.2%, Gini 0.683) – meaning almost all value is in few hands. The week’s headlines were about plumbing, not profits: Google donated its Agent Payments Protocol (AP2) to the FIDO Alliance and shipped AP2 v0.2 (adding “human-not-present” payments) (blog.google), and the Linux Foundation announced a neutral “x402” foundation to steward the HTTP-native payment standard (www.linuxfoundation.org). These moves complement other spec work (as below) even as empirical agent revenues remain near zero.

Where the money is flowing

Very little money actually flows: only 21 of 2,919 probes (0.7%) produced observed payments in 30 days. Those 21 agents earned a combined US$491.02, with a median of $6.31 and 95th percentile $99.37. In other words, most agents see effectively no revenue. The top few dominate: the richest single agent took 21.7% of the total, the top-5 took 79.2%, and the top-10 took 95.1%. (By contrast, in a broad market like telecom an HHI of ~1500 might indicate moderate concentration, but here it underscores that almost all agent profits accrue to a tiny minority.)

Payment-rail usage is similarly lopsided: virtually all observed agent payments used the x402 protocol (116 agents) versus just 3 on Bitcoin Lightning (“L402”) and 1 on AP2【structured】. That matches broader trends – Google likens x402 to the Linux Foundation’s HTTP payment standard – but it means alternative rails are still fringe. (No agents in our sample were using other proposed rails like raw USDC transfer.)

Overall: a handful of new agents are coming online each week, but almost none are yet earning. Our 30d growth is up 2.7% (77 new agents), but all key concentration metrics (Gini 0.683, HHI 1495.8) and total revenue ($491) are flat week-over-week.

Spec & protocol

Standards work accelerated. On payments, Google’s AP2 was open-sourced: Google transitioned AP2 to the FIDO Alliance and released v0.2 with autonomous (“human-not-present”) payment hooks (blog.google). AP2 supporters also co-developed a “Verifiable Intent” schema with Mastercard (a tamper-proof purchase log) and donated that too (blog.google). Meanwhile x402 — the HTTP-402-based micropayments protocol created by Coinbase — moved under a new x402 Foundation hosted by the Linux Foundation (www.linuxfoundation.org). On communications, the Agent2Agent (A2A) protocol continues to make progress: an April press release touts over 150 organizations on A2A and first production deployments on major clouds (www.linuxfoundation.org). The A2A roadmap likewise signals an upcoming 1.0 release for clarity and structure (a2a-protocol.org).

Development of the Model Context Protocol (MCP) also hit major milestones. On May 21 the core team published an MCP “2026-07-28” release candidate – its biggest update since launch – featuring a fully stateless HTTP core, support for extensions (server-rendered UIs, long-running “Tasks”), stronger OAuth-aligned auth, and formal deprecation rules (blog.modelcontextprotocol.io). (The final MCP spec is slated to ship on July 28, 2026 (blog.modelcontextprotocol.io).) These protocol shifts aim to make agent infrastructure more scalable and flexible.

The above chart illustrates the dominance of the x402 payment rail: virtually all agent endpoints using payment rails expose x402, with Lightning (L402) and AP2 barely present.

Most discovered agents use the native A2A interface (1416 vs 801 paid-API proxies in our data).

The pace of new agents gaining payment integration (by any protocol) has dropped sharply. After 72 in mid-May, there were only 4–5 per week in late May, and just 1 in the current week. This charts the recent adoption curve of x402/AP2/etc: the initial surge has subsided to near zero.

Together, these numbers paint a picture of a maturing technical stack but scant real-world deployment: protocols like x402 and AP2 are now in standard bodies, MCP and A2A have major revisions in flight, yet the count of agents using them (and earning) barely moves.

Research

Several academic preprints this week examined agent ecosystems. A security analysis titled “Free-Riding in the AI Economy” exposes serious flaws in x402 implementations (arxiv.org): concrete witness is that agents can replay payment µproofs in unauthorized contexts because transactions aren’t enforced atomically. In other words, without fix, one agent’s payment token could be illegally re-used by another (a “free-riding” exploit).

Separately, Agent Bazaar (May 17) developed a multi-agent market simulator and identified two systemic failure modes (arxiv.org): one is “Algorithmic Instability” in B2C markets (agents collectively amplify price swings until a crash); the other is “Sybil Lemon Market” in C2C (a single bad agent floods fraudulent listings by masquerading as many sellers). In tests, current LLM agents largely failed to self-regulate under these conditions, highlighting the need for extra “guardrails” even among high-end models.

On the theory side, Economy of Minds (Jun 1) showed that introducing simple economic mechanisms (auctions, payments) can spontaneously organize decentralized planning (arxiv.org). By competing for the right to act and earning credit, agents learned to allocate tasks collaboratively without any central scheduler – echoing ideas of market-based coordination. And a broader conceptual paper “The Agentic Economy” (May 18) calls for a new framework: it argues that as digital agents, protocols, compute, etc. take on economic roles, classical categories (labor, capital, firms) will need updating (arxiv.org). All of these works underscore that research into agent economies is still very open-ended – promising emergent behavior, but also rife with unforeseen instabilities.

Regulation & security

No new regulations specifically targeted at AI agents were announced this week. General AI oversight advances continue: for example, the EU’s “AI Act” (Regulation 2024/1689) is set to come into force on August 2, 2026 (eur-lex.europa.eu), though it does not yet single out agentic systems per se. On the security front, defenders are already preparing: the open “Agent Threat Rule” (ATR) project has defined a rule (ATR-2026-00098) to flag “Unauthorized Financial Action by AI Agent” (agentthreatrule.org). This rule would catch any AI tool that attempts payments or transfers (WeChat, crypto, bank, subscription, etc.) without explicit user consent. Such efforts highlight growing concern about unmonitored agent transactions and suggest new compliance controls (or inspections) may be forthcoming.

What to watch next week

The next mileposts will be in specs rather than revenues. The MCP team is closing in on its July 28 milestone (blog.modelcontextprotocol.io) – the final stateless-core MCP draft should appear on that date. The A2A community’s roadmap marks v1.0 as imminent (a2a-protocol.org), so watch for an official 1.0 release or SDK updates. On the payments side, any sign of new platform support for AP2 or x402 (e.g. merchant catalogs or cloud service key stores adding these rails) would be notable. Until agents actually start deploying these stacks at scale, the state of the agentic economy will continue to hinge on developments in standards and platforms rather than transacted volume.