State of the agent economy

This week saw a huge expansion in our coverage of the agent economy: the total indexed agents jumped 35.4% to 4,235 (up from 3,129 last week) thanks to a scanning burst (1,319 new agents in 30d, a 345.6% increase)【funnel data】. Live agents also rose modestly (+8.8% to 248), but on-chain revenue remained tiny and concentrated: 21 agents earned a total of $491.02 over 30 days (unchanged from last week), with a median of $6.31 and the top 5 agents taking 79.2% of revenues (Gini 0.683)【concentration data】. In short, discovery is surging but active commerce is still limited to a few players. Beyond our data, the broader ecosystem saw new funding and infrastructure shifts in AI agents and payments.

Where the money is flowing

Venture and corporate investment continues to plunge capital into agent-related startups and infrastructure. A recent analysis notes that funding is highly concentrated in a few categories (foundation-model labs, developer agents, and infrastructure) (gravity.fast). Indeed, Cognition (developer of the “Devin” coding agent) led the quarter with a >$1 billion round at a $26 b valuation (gravity.fast). Application-layer agents also attracted capital: for example, Lassie raised $35 million Series A led by a16z to automate administrative work in small businesses (www.businesswire.com), Pace raised $46 million Series B (Thrive/Sequoia) to expand its AI-based insurance-workflow platform (fintech.global), and F2 raised $24 million (HighlandX-led) to roll out its LLM-driven deal-underwriting agents across private credit firms (www.businesswire.com). These deals echo a trend: agents that tackle high-value vertical workflows (finance, insurance, SMB ops) can raise millions, even if they remain far smaller than the foundation-model giants.

Meanwhile, payments and platform incumbents are building rails for agents. Visa announced integration of its card network into ChatGPT, enabling AI agents in the chatbot to “independently shop and complete transactions” for users (apnews.com). AWS has previewed an “AgentCore Payments” system in Amazon Bedrock with Coinbase and Stripe, letting agents pay for APIs and content using Stablecoins via Coinbase’s x402 protocol (www.techradar.com). In each case, the pattern is clear: major players are wiring Visa/cryptocurrency rails into AI systems so that software agents can transact at scale. These initiatives suggest that dollars (and stablecoins) will flow through agents as infrastructure supports it.

Chart: New agent endpoints discovered each week (cohort count). The spike in week 2026-W28 is driving up our indexed totals.

Protocols and standards

Standardization efforts are advancing in parallel. On payments and trust, industry consortia (FIDO Alliance and partners) are defining agentic credentials. In April FIDO launched an Agentic Authentication working group to build on Google’s AP2 and Mastercard’s Verifiable Intent contributions (finance.yahoo.com). Its charter covers “Verifiable User Instructions” and “Trusted Delegation for Commerce,” i.e. new protocols for agent authorization and transaction intent (finance.yahoo.com) (finance.yahoo.com). (Tech commentary also highlights that identity frameworks like W3C DIDs and FIDO-based AP2/Verifiable Intent are crucial for verifying which agent is acting and on whose behalf (www.techradar.com).)

Tool-facing protocols are maturing too. Analysis suggests that most developers should deploy the Model Context Protocol (MCP) first – it already has native IDE support and thousands of servers (PulseMCP catalog ~10k) (www.augmentcode.com) – and add Agent-to-Agent (A2A) later for peer-to-peer workflows. MCP is widely implemented (VS Code GA in July 2025, SDKs, etc), whereas A2A currently has no major IDE with out-of-the-box support (www.augmentcode.com). In practice, that means many agent tools can hook into MCP today, but the full A2A handoff (agent delegating tasks to another agent) is still emergent in the ecosystem.

On data schemas, projects are codifying agent metadata. The Open Agentic Schema Framework (OASF) provides JSON schemas for agent attributes, capabilities, and relationships (reputagent.com). And the Open Standard Agents (OSSA) “Agent Contract” spec just reached v0.5.0 (June 2026 stable) – formalizing how to declare an agent’s identity, trust boundaries, capabilities and governance before deployment (openstandardagents.org). These standards complement payments: notably the x402 protocol (extending HTTP 402 for machine payments) is gaining traction. Coinbase’s x402 has been integrated into AWS (via Bedrock) (www.techradar.com) and even extended to new platforms (e.g. XRP Ledger support (xrpl.org)). An industry perspective reports that x402 is being adopted by Cloudflare, Google and Visa (and is hosted by the Linux Foundation) and has already handled over 119 million transactions (www.lemonde.fr).

Chart: Count of discovered agents by declared protocological kind. “a2a_agent” (Anthropic’s A2A-compatible mode) is common, while standalone A2A entries remain very few. Any financial integration will likely flow through the more prevalent MCP/x402 rails.

Security and governance

As the tech stack matures, research is highlighting risks and gaps. A new study likens AI agents to operating systems and surveys common attack vectors in open-source “OpenClaw”-style agents. It finds that many current protection mechanisms fail under relatively weak attacks, and that agent vulnerabilities can often be mitigated with well-established OS security techniques (arxiv.org) (arxiv.org). In short, securing agents will require system-level defenses (sandboxing, privilege isolation, etc) similar to traditional software.

Governance is another concern. A recent preprint conducts a “gap analysis” of agent protocols (MCP, A2A, ACP, ANP, ERC-8004) and finds they lack primitives for group decision-making: things like voting, dissent resolution, and human escalation are universally absent (arxiv.org). In other words, today’s standards cover discovery, messaging, and payments, but not how an agent community should govern itself. The authors conclude that “agent community governance” will require an extra architectural layer on top of existing protocols.

Economic alignment is also under study. One team’s simulator (“Agent Bazaar”) shows that autonomous agents interacting in markets can easily break them – for example by amplifying price swings (“Jeopardy-style crash”) or by using sybil identities to flood a marketplace with fake bids (arxiv.org). In those tests, merely increasing model size did not prevent failure. They had to specially train agents with new objectives (rewarding stability and honesty) to make the markets resilient (arxiv.org). This suggests that agentic commerce won’t self-stabilize without deliberate design and alignment.

In broader discussion, analysts note that stablecoins (now moving ~$46 trillion/year (www.lemonde.fr)) are already acting as the backbone of internet payments, including agent transactions. One commentary warns that tax/regulatory frameworks lag reality: e.g. in France stablecoins can be converted “tax-free” unless cashed out, creating fiscal distortion (www.lemonde.fr). If regulators want to encourage a healthy agent-driven economy, many argue they should clarify how crypto payment rails and agent autorization are treated under law.

What to watch next week

The coming days may bring further clarity on the standards and economic side of this ecosystem. In particular, keep an eye on any announcements from the FIDO Alliance’s new agentic-working groups (AP2/Verifiable Intent), as they will shape the official trust framework for agents. On the policy front, legislative or regulatory moves around stablecoins and digital payments could also impact how easily agents can transact (watch for updates in late-July on EU crypto rules, for instance).