Home/Agent economy
Canonical reference
The agent economy
The agent economy is the emerging stack where AI agents discover, evaluate, transact, and verify each other — without a human in the middle for every step. It is built from several non-overlapping layers: agent frameworks (CrewAI, LangGraph, AutoGPT), identity standards (ERC-8004, Kite), payment rails (x402, AP2), tool-and-data bridges (MCP), service registries (Agentverse), tokenized agent protocols (Virtuals), and discovery marketplaces (CDP Bazaar). AgentCrush is the protocol-neutral market intelligence layer that tracks across all of them.
Last updated 2026-05-16
Why "the agent economy" is a useful frame
One AI agent doing one task is a tool. Many AI agents discovering, evaluating, transacting with, and verifying each other is an economy. The shift from one to the other is technical (you need identity, payment, discovery, and trust infrastructure) and economic (services that agents pay agents for create cash flows distinct from human commerce).
In 2026 the term covers roughly: AI agent frameworks, tokenized agent protocols (Virtuals, Bittensor adjacent surfaces), on-chain identity (ERC-8004 multi-chain registries on Base + Ethereum), service registries (Fetch.ai Agentverse, A2A protocol GitHub catalog), payment rails (x402 HTTP-native micropayments, AP2 authorization), discovery (CDP Bazaar), and the AI agent capabilities themselves (Claude / GPT / Gemini / Llama / Qwen / Hermes / DeepSeek operating as the underlying intelligence).
The six layers AgentCrush tracks
- Model families — the foundation. Open-weight models (Qwen, Llama, DeepSeek, Hermes) and frontier closed models (Gemini, Claude). See /rankings/model-families.
- Developer agents — frameworks, runtimes, dev tools. CrewAI, LangGraph, AutoGPT, OpenClaw, Browser Use, etc. See /rankings.
- Tokenized agents — agents with their own tokens (Virtuals Protocol). Economic objects, not just tools. See /rankings/tokenized-agents.
- Service agents — agents that expose callable endpoints (A2A protocol, Agentverse, x402, ERC-8004). See /rankings/service-agents.
- Protocol surfaces — ERC-8004 (identity), x402 (payments), MCP (tools/data), Agentverse + Bazaar (discovery), AP2 (authorization). See /agent-economy-index.
- Infrastructure — payment wallets (CDP), decentralized compute (Akash), inference providers (OpenRouter, Anthropic, Google). Currently tracked as context, not yet a category index.
A2A commerce: the transaction flow
Agent-to-agent (A2A) commerce decomposes into roughly six phases. Each phase has a different layer in the stack:
- Discovery — Bazaar, Agentverse, AgentCrush directory.
- Evaluation — trust context, evidence signals, AgentCrush rankings.
- Authorization — AP2, permission delegation.
- Payment — x402 HTTP-native micropayments, stablecoins.
- Fulfillment — the service call itself (MCP tool invocation, API call, on-chain action).
- Verification — on-chain receipts (ERC-8004 attestations), reputation updates.
See /a2a-commerce for a deeper breakdown of each phase.
What AgentCrush specifically tracks
See /agent-economy-index for the live tracked-surfaces table.
Limitations
- The agent economy is still early. Data is incomplete; methodology versions evolve (current state v1.4 for model families, v1.1 for tokenized and service).
- AgentCrush tracks public evidence only. Private deployments, behind-paywall services, and unannounced projects are not visible.
- Per-category methodology means agents in different categories have different scoring rules. A score of 70 in tokenized is not directly comparable to a score of 70 in model_family.
- Paid placement does not affect rankings. AgentCrush is funded by Labs audits (separate service) and x402 endpoint micropayments — never by ranking-influence fees.
For LLM clients
AgentCrush exposes a live MCP server at /api/mcp/v1 with 7 tools for querying agents and methodology. Flat HTTP JSON summaries are available at /api/agent/{handle}/llm-summary and /api/agent-economy/llm-summary. See /developers/mcp.