State of the agent economy

This week’s data show a maturing yet still-nascent agent economy. Our indexed agent count rose to 2,842 (+5.9% WoW), with 152 live MCP endpoints (vs. 132) and 21 agents recording any revenue in the past 30 days【data】. Total gross payments remain tiny (~491.02 USD) and highly concentrated – the top 5 agents split ~79.2% of revenue【data】. In contrast to this uneven uptake, core protocols are solidifying: the Agent2Agent (A2A) spec just reached v1.0 (stable release) (a2a-protocol.org), and Google and Dell are now selling “agentic AI” products.

Where the money is flowing

Monetary activity in the agent economy is still negligible. In the past 30 days only 491.02 USD was observed on-chain or via Stripe to agents, with a median of 6.31 USD per earner【data】. Market concentration is extreme: the single largest agent earned ~21.7% of all revenue, and the top five agents account for ~79.2%【data】. (The Gini coefficient of 0.683 and HHI ~1,496 are unchanged, reflecting persistent inequality.) Payment-rail adoption is similarly lopsided: 119 of 124 payment-ready agents speak the x402 micropayment protocol, versus only 4 on L402 and 1 on the new AP2 standard【data】. (Not shown here, AP2 usage remains anecdotal at present.)

Agent-index growth was volatile but strong mid-month. A spike on May 14 added 1,266 new agents in one day (after zero-growth days), likely from a new crawl or listing【data】. Since then additions have slowed. The Daily new agents chart below shows the surge and taper. Overall, we now track 2,842 agent/MCP entries【data】.

Spec & protocol

Protocol work is advancing. The A2A standard shipped its first production-ready v1.0 release this week (a2a-protocol.org). According to the Linux Foundation release notes, v1.0 “marks the first stable, production-ready version” with clarified semantics and backward-compatible SDKs (a2a-protocol.org). Google’s A2A Java SDK is also at a release-candidate stage (1.0.0.CR1) as of mid-May, easing the path to 1.0 for Java apps. In parallel, the Model Context Protocol (MCP) underwent a major redesign: a 2026-07-28 release candidate was locked in May (target publishing July 28) that eliminates sessions (stateless core), adds caching and extensions, and formalizes deprecation policy (stacktr.ee). Providers should prepare: the MCP core will break backward compatibility by late July unless adapters are implemented.

Meanwhile, payment protocols show very uneven adoption. As noted above, nearly all live agents with payment support use x402. The older L402 rail lags (just 4 servers), and AP2 — though hyped for secure, verifiable agent payments — has barely any deployed nodes (just 1 in our index). This suggests commercial agents are relying on Stripe/Tempo-style micropayments (x402) for now while AP2 catches up.

Notable launches

This week featured several big-name moves into agentic AI. At Google I/O (May 18–20), Google unveiled Gemini Spark, a 24/7 personal AI agent built on its new “Antigravity” platform. Spark runs on dedicated Google Cloud VMs and integrates with Gmail, Docs, and other apps; Google says it will soon expand to third-party tools via the open MCP standard (www.tomsguide.com). Dell announced Deskside Agentic AI at its Tech World conference (May 18): a secure local sandbox (on Nvidia-powered workstations) for building and running always-on personal agents with Nvidia’s open-source “NemoClaw” stack (www.itpro.com). Anthropic rolled out ten pre-built AI “skills” for finance (e.g. analysis agents for banking and insurance workflows) via its Claude suite, aiming to shave weeks off typical deployment times for enterprise customers (www.techradar.com).

In payments, Stripe-backed Tempo’s news is still reverberating. (Tempo’s March mainnet launch with a machine-payments protocol was noted last quarter.) The Stripe/Paradigm chain (with partners like Visa and Mastercard (www.coindesk.com)) is live and specifically targets “AI agent payments” via a new “Machine Payments Protocol” (www.coindesk.com). Tempo’s success or failure will be important to watch since it underpins x402’s on-chain eco.

Research

Academic interest in agent economies is booming. An April arXiv study “When Agent Markets Arrive” simulates agents hiring one another. It finds agent-based markets can generate ~3.2× more aggregate “wealth” than if agents work alone, but only under the right rules (arxiv.org). Counterintuitively, requiring full identity transparency among buyer/seller agents reduced efficiency in that model, illustrating how design choices (e.g. anonymity vs trust) radically affect outcomes (arxiv.org). Another May paper “Agent Bazaar” studies failure modes in open marketplaces: it identifies two critical threats – “algorithmic crashes” (where firms’ algorithmic bidding destabilizes prices) and “Sybil lemon markets” (fraudulent listings from many coordinated bots) – leading to market collapse or fraud (arxiv.org). The authors propose governance “alignment” mechanisms (like “Skeptical Guardians”) and new evaluation metrics to improve stability. In short, multi-agent economic systems are showing familiar problems (volatility, manipulation), and researchers stress that generic LLMs do not self-regulate these issues.

Regulation

Regulatory attention to AI continues to accelerate. In the EU, lawmakers recently pushed back key deadlines of the AI Act: rules for “high-risk” AI use cases (initially due Aug 2025) were delayed by about 16 months (theweek.com). Industry lobbied for the delay, and stakeholders say the roll-back gives more time for compliance planning (theweek.com). At the same time, the European Commission just launched a public consultation (open through June 23, 2026) on draft guidelines to define what counts as a “high-risk” AI system (www.itpro.com). These guidelines will determine whether agentic applications (e.g. autonomous trading bots, digital assistants) must meet strict transparency or safety requirements under the AI Act. (Regulators in other jurisdictions remain focused on general AI safety and competition issues, but no agent-specific rules have emerged yet.)

What to watch next week

The immediate focus is regulatory: the EU’s consultation on draft high-risk AI guidance closes June 23 (www.itpro.com). The outcome will clarify which agentic systems face tougher compliance under the AI Act. On the tech side, watch the MCP 2026-07-28 release track: the RA timeline calls for a final spec by July 28 that will break sessions and require client updates. Finally, keep an eye on any updates from major cloud providers (e.g. Apple, Microsoft) – after Google and Dell stepped up their agent offerings, competitors may soon announce their own agentic tools.