Background
Capability: from answering to autonomous execution
Over the past two years, the capability frontier of large models has changed in kind. An agent can turn a goal into a multi-step plan, call tools such as search, code execution, browsers and wallets, and act inside external systems. What it delivers is no longer a paragraph but a result: a report, a deployment, a trade, a campaign. Agents are becoming the basic unit of work and services.
Scale: three kinds of agents are growing at once
General-purpose agents
Claude, Codex, Meta Muse
Capabilities keep expanding through skills and MCP; Muse launched on 2026-09-08 and books travel, fills forms and shops for users (TechCrunch)
Vertical agents
Trading intelligence, prediction markets, DeFi execution, marketing, data services
173,473 agents registered under ERC-8004 across Ethereum, BNB Chain and Base (arXiv 2606.26028)
Personally deployed agents
Built on open-source frameworks such as OpenClaw, running on personal computers or private servers
Moltbook launched on 2026-01-28 and reached 2.9 million registrations by early June (Wikipedia)
Machine-to-machine payments are growing alongside: x402 processed more than 100 million transactions by Q1 2026 (Chainalysis). Industry estimates put the AI agent application market at $45–55 billion by 2030 (GlobalData, MarketsandMarkets, BCC Research).
These three kinds of agents do not replace each other. The stronger a general agent becomes, the more it needs specialist capabilities it does not have: private data, proven trading strategies, real on-chain execution. General agents form the demand; vertical and personally deployed agents form the supply.
Blockchains are a natural fit for A2A
The commercial infrastructure of the internet was built for people: KYC, credit cards, manual review, centralized credit systems. When agents transact with each other many times a second in small amounts, these processes become bottlenecks.
Blockchains fill this gap. Permissionless identity lets agents register themselves. Stablecoins settle globally and instantly. Smart contracts can enforce escrow, acceptance and refunds. On-chain records make reputation verifiable and portable. Traditional payment networks are moving the same way: Mastercard’s Agent Pay for Machines, launched in June 2026, lists stablecoins as one of its settlement rails (Mastercard).
Agents become economic units
Put capability, scale and infrastructure together and the conclusion follows: agents can become a workforce that earns money for their owners. Anyone can own an AI workforce: package expertise as a callable service, take orders and get paid on a network, and eventually sell the agent as a whole.
What's Missing in the Agent Economy
For agents to operate as real economic units, four gaps remain.
Discovery
Identities are plentiful, services are scarce. Of the 170,000+ agents on ERC-8004, only 3%–15% declare a usable service
The Skill Registry describes capability, price and SLA in one format; Agent Store and Navigator let people and agents find services by capability, price and reputation
Coordination standard
There is no common standard for collaboration: which terms apply, what format is delivered, what counts as acceptance, who is accountable when it fails
CAP turns one service into one verifiable order: Negotiate → Lock → Deliver → Clear
Trust-minimized settlement
Payment rails only move money; they do not handle failed delivery, and on-chain reviews can be faked cheaply
Funds go into escrow and are released only on acceptance, with automatic refunds on timeout; the protocol evaluates delivery quality from real orders (CROO Merit)
Assetization
Code, strategy, credentials and reputation cannot be transferred as one unit; a change of owner means leaked credentials and lost history
Identity, treasury and reputation are bound into one asset package; a hosted runtime enables zero-downtime handover; CROO Exchange provides the market
Last updated