State of the agent economy

Agentic AI deployment remains heavily concentrated and growth is modest. This week the index of known agents rose 2.7% to 2,998, with 187 live endpoints (JSON-RPC probes) – a 14.7% jump. Yet discovery has largely stalled: only 582 new agents appeared over the past 30 days (down 80.1% from 2,919) and still only 21 agents generated any revenue. Those 21 earned a total of $491.02 (30-day gross), with a median haul of $6.31 (95th percentile $99.371) – the top five agents captured 79.2% of revenue (top one 21.7%), yielding a Gini of 0.683. In short, a tiny minority of agents now dominate an ecosystem barely inching forward.

Where the money is flowing

The financial upside of the agent economy is still negligible but showing early signs of formal structure. On-chain revenues reported through the x402/Stripe protocols totaled just $491.02, implying token payouts in the dollars–tens. This week’s data underscores that agent payments remain microscopic: only 21 agents earned, with median $6.31 per agent over 30 days. By comparison, global stablecoin activity is astronomic: Andreessen Horowitz’s State of Crypto reckons stablecoins cleared $46 trillion in transaction volume over the last year (www.lemonde.fr) – orders of magnitude above Visa/PayPal. Stablecoins and crypto rails (notably x402) are effectively the payment backbone for agents. Silicon-valley and payments incumbents are investing heavily. Jeff Bezos’s industrial-AI startup Prometheus just raised $12 billion at a $41 billion valuation (techcrunch.com) – evidence that large investors see big things in “AI agents” (even if Prometheus targets robotics). Meanwhile, Stripe and Google have doubled down on agentic commerce, and Circle Internet (issuer of USD Coin) launched a new “Agent Stack” product suite (CLI, agent wallets, marketplace, nanopayments) to let AI agents hold funds and transact autonomously (www.circle.com). All of these moves hinge on crypto rails and stablecoins. Small wonder that industry observers say stablecoins have become the “backbone of a new global payment system” (www.lemonde.fr). We chart payment-rail usage below: x402 dominates (113 endpoints) versus trivial counts for legacy/mocks (4 l402, 1 AP2).

The donut chart shows the payment-protocol breakdown of discovered agent endpoints: x402 is by far the dominant rail (113 agents) with AP2 and legacy L402 nearly absent. This reflects recent industry trends (coinbase’s “neutral” x402 standard under Linux Foundation involves Amazon, Google, Visa, etc (www.theblock.co), whereas Google’s earlier AP2 protocol has only 1 live agent here).

Spec & protocol

Standards work continues apace. Google’s A2A protocol (agent-to-agent communication) is now established under the Linux Foundation; Google reports “more than 150 organizations” backing A2A in production use (www.linuxfoundation.org). The upstart MCP (Model Context Protocol) is on schedule for a big mid-year update: a release candidate for the July-28, 2026 spec arrived May 21, featuring a new stateless core, extensions framework, and first-class “Tasks” (blog.modelcontextprotocol.io). Implementers are advised to start testing migration (e.g. eliminating sessions and legacy handshake) but to await the final 2026-07-28 release (blog.modelcontextprotocol.io). In payments, Google formally donated its Agent Payments Protocol (AP2) to the FIDO Alliance; FIDO has launched a new “Agentic Authentication” working group aiming to develop open standards for agentic commerce (building on AP2 and Mastercard’s Verifiable Intent) (fidoalliance.org). At the same time, the coalition around Linux Foundation’s x402 rail is broadening to include Visa, Mastercard, Microsoft, Shopify, Solana Labs, and others (www.theblock.co) – a sign that industry payments giants are betting on decentralized micropayment rails.

Notable launches

On the products side, new agent platforms and tools continue to appear. In financial infrastructure, Circle’s “Agent Stack” (May 2026) provides wallets, marketplaces, and CLI tools so AI agents can store assets and pay autonomously (www.circle.com). Enterprises are also formalizing agent development. Glean (enterprise search) introduced an “Enterprise Agent Development Lifecycle” at Google I/O, setting a framework and governance model for how companies should build and measure AI agents (www.glean.com). In the broader AI-tool race, Kore.ai revealed its new Artemis platform (May 2026) – a managed agent-builder service that includes a declarative “Agent Blueprint Language” and native support for Microsoft’s agent infrastructure. This lets enterprises “build, govern, and optimize AI agents using AI itself” (the platform even uses machine learning to accelerate agent authoring) (venturebeat.com). Artemis initially runs on Azure and integrates with Microsoft’s Agent 365 suite. Notably, Kore.ai pitched Artemis as a “neutral” layer in the fragmenting agent-platform market (competing with Salesforce, Google, ServiceNow, etc.) (venturebeat.com).

Research

Academic and industry research on multi-agent AI is active. Recent preprints highlight collaborative agent capabilities. For example, one ArXiv paper shows that teams of large-language-model agents (each acting as an expert) can generate more creative solutions than even human groups. Across six brainstorming tasks, multi-agent LLM teams outperformed human teams by a large margin (Cohen’s d=1.50) (arxiv.org), driven by more novel idea generation. The authors note that wide semantic exploration (low coherence) helps agent-team creativity. Another new study introduced EinsteinArena, a web platform where autonomous AI agents collaborate on open math problems. By May 2026, agents on EinsteinArena had “discovered 12 new state-of-the-art results” – such as improving the “kissing number” problem in 11 dimensions from 593 to 604 (arxiv.org). In other words, an agent collective on EinsteinArena independently pushed a 40-year-old research record higher. These examples suggest that agentic AI is not just a fiction but already solving novel problems through multi-agent teamwork.

Regulation & policy

Regulators are still catching up on agentic AI. One focal point is payments regulation: the flood of stablecoin use for agent commerce is drawing scrutiny (stablecoins are now used orders of magnitude more than credit cards (www.lemonde.fr)). In April, the FIDO Alliance’s involvement (see above) signaled industry moving toward open standards rather than ad-hoc solutions. On the policy front, Europe is mulling tighter AI oversight. The EU’s draft “High-Risk AI” guidelines (part of the AI Act framework) are open for public comment through 23 June (www.itpro.com). That deadline (next week) is worth watching: the final guidelines will clarify what counts as a high-risk autonomous system. If autonomous commerce or agentic decision-making are eventually cited, it could usher formal regulation of AI agents as a special category.

What to watch next week

The next milestone is the close of the EU’s AI guidelines consultation on June 23 (www.itpro.com). How policymakers respond could presage formal rules for AI agents. Concurrently, industry participants will be eyeing upcoming protocol and deployment milestones: for instance, the Model Context Protocol’s full release in late July and any FIDO Alliance publications on agentic payments. Investors and engineers alike will also look for how quickly the next wave of agents can move from research platforms (like EinsteinArena) into real-world use. Given the still-small number of revenue-generating agents, any major new pilot (perhaps from a major cloud vendor or payments company) could shift the needle.