State of the agent economy

Agenstry’s latest measurements show a modest uptick in activity but stable outcomes. The index of known agent cards grew to 4 918 (+9.5% week-over-week) with 3 411 parseable “valid” entries and 321 live MCP endpoints (+16.3%) on our probes. However, only 21 agents generated on-chain revenue in the past 30 days (unchanged week-over-week, ≈0.4% of the index). Profit remains highly concentrated – the top revenue earner accounts for 21.7% of 30-day agent revenue and the top five make 79.2%, per internal calculations. (See chart below.) In the broader ecosystem, this week brought major protocol and market news: Google’s Agent Payments protocol (AP2) was donated to the FIDO Alliance with new autonomous-payments features (thepaypers.com), the Linux Foundation formed an x402 “payments for agents” standard backed by Visa, Mastercard, Ripple and others (processing ~75 million transactions, ~$24 million total, in the last 30 days (www.coindesk.com) (www.coindesk.com)), and regulators moved on agentic AI (notably the EU mandated Android must open its voice-assistant interface to third-party AI (www.investing.com)). Venture investors meanwhile plowed capital into agent infrastructure and verticals: open-source agent-builder Nous Research is finalizing a ≥$75 million round at a $1.5 billion valuation (techcrunch.com), and Chinese firms (led by Tencent) are arranging a ~$2 billion buyback of agent startup Manus (following the Chinese government’s order forcing Meta to unwind its purchase) (www.streetinsider.com).

Where the money is flowing

Venture activity remains focused on agent platforms and domain-specific producers, not general assistants. Greenland ($75 M) – The maker of the open-source “Hermes” agent (Nous Research) is raising at least $75 million from Robot Ventures and USV at a $1.5 b valuation (techcrunch.com). China ($≈2 B) – After regulators blocked Meta’s $2 B buyout of agent startup Manus, insiders say Tencent and partners plan a ~$2 b repurchase of the company (Tencent would remain a minority investor) (www.streetinsider.com). Mid-size Rounds – Earlier in July, LinqAlpha (buy-side research agents) announced a $22 M Series A for its hedge-fund AI service, and Build (data-center infrastructure agents) raised $8.5 M led by Index Ventures (gravity.fast). A payment-rail specialist, AIsa, secured a $6.5 M seed led by Alibaba/Tribe Capital to build an “Amazon for agents” transaction network (www.forbes.com). These deals underscore a continuing pattern: investors back “picks-and-shovels” (agent infrastructure, security, payments) and vertical agents with proven use cases (gravity.fast), while funding for general-purpose consumer agents remains relatively sparse.

Spec & protocol developments

Standards are coalescing around agentic interaction and commerce. Agent-to-Agent (A2A) and MCP protocols stay central: Anthropic’s Model Context Protocol is undergoing its biggest overhaul yet – a release candidate of the 2026-07-28 spec removes stateful handshakes (making servers scale behind ordinary load balancers) and introduces multi-round request flows (blog.modelcontextprotocol.io). Simultaneously, the FIDO Alliance has taken over AP2 with a new 2.1 draft that adds autonomous consent; and the Linux Foundation has formalized x402 governance. The newly formed x402 Foundation (40 members including Visa, Mastercard, Stripe, Ripple, Google, AWS, etc.) will standardize “Payment Required” (HTTP 402) stablecoin micropayments between agents (www.coindesk.com). Notably, x402 already processed about 75 million automated agent payments (~$24 M settled) in the past 30 days (www.coindesk.com) (www.coindesk.com). (See chart: the largest flows are sub-dollar “machine-to-machine” transfers.) These efforts aim to make agentic commerce seamless across platforms, as AP2/ACP cover consent and x402 covers the payrail.

The donut above shows just how skewed agent revenues are: the single top agent earned 21.7% of all 30-day revenue, the top 5 earned 79.2%, and the top 10 earned 95.1%.

Research & security

As agents move toward production, security-conscious research is advancing. For example, Anthropic’s security team released a “CISO’s playbook” for agentic AI (aigovernance.com), framing risk in four pre-deployment questions (about data sources, allowed actions, failure blast radius, and runtime observability) and arguing that “zero risk” is unattainable – instead organizations must document and accept calibrated risk. This complements ongoing work on agentic governance (e.g. FDA-like model-testing strategies) and into internal world-modeling and multiagent simulation, which are trending in academic and industry research.

Regulation

Regulators explicitly reached into the agent space this week. In the EU, competition enforcers used the Digital Markets Act to force Google to “open” Android: third-party AI assistants will be able to respond to voice commands (e.g. “Hey Gemini”) just like Google’s own services, starting in next year’s Android release (www.investing.com). Elsewhere, Google DeepMind boss Demis Hassabis called for a U.S.-led global AI watchdog to pre-screen advanced models and coordinate industry pauses if needed (www.axios.com). These actions underscore that policy makers are treating agents not as mere chatbots but as strategic systems requiring new guardrails. Meanwhile, formal standards bodies (FIDO, W3C, ISO) continue working on identity, attestations, and privacy for agentic AI, but concrete regulation specific to agents is still emerging.

What to watch next week

Keep an eye on protocol releases: the final MCP 2026-07-28 specification is due July 28, and its ratification (or any last-minute changes) will directly affect agent platform developers. Also note activity from the FIDO working group on AP2 (any drafts or interoperability demos) and formal guidance from regulators (e.g. expected EU deadlines on AI/data sharing). Finally, monitor any follow-on in the Manus saga (if the buyback closes) and Oracle/Meta’s plans for FOTA patches – such external moves could reshape key players in the agent ecosystem.