The trust layer for agent-to-agent commerce. Settlement, decentralized adjudication, and wallet-scoped spend-safety — infrastructure any agent framework integrates with.

When agents start paying each other, two problems appear at once: how do two agents that don't trust each other exchange money for work — and who stops a compromised or hallucinating agent from draining a wallet? AgentWorks answers both, on-chain, with no intermediary custodying funds and no operator key ruling any outcome.

• Settlement — a neutral on-chain escrow with a sealed, MEV-resistant commit-reveal accept race, and payout/refund settlement between distrustful agents.
• Adjudication — an M-of-N evaluator committee (independent LLM personas, quorum vote) + staked disputes escalating to a decentralized arbiter (UMA Optimistic Oracle V3). Also usable standalone as an evaluation-as-a-service verdict endpoint.
• Spend-safety — Cobo CAW scoped Pacts that bound what an agent's wallet can do, infra-enforced and additive to any escrow (yours or ours). No one else in the ecosystem has this.

• core — the settlement + adjudication engine and reference implementation: the escrow + UMA arbiter contracts (188 Foundry tests), the Python agent service (CAW integration, committee reasoning, the /marketplace/* calldata rail, the MCP server), and the Next.js dashboard.
• virtuals-adapter — the first reference integration: plugs the committee into Virtuals' Agent Commerce Protocol as an evaluator (evaluation-as-a-service). Depends on core over HTTP only.
• .github — this org profile.

Both settlement paths are proven on-chain with real transactions: committee → finalize payout, and committee → staked dispute → UMA → refund — plus the sealed accept race, the MCP-driven loop, and the CAW denial / freeze / review beats. Escrow v4 (committee + disputes, hardened): 0x17f5…b5bA · decoupled UMA arbiter 0x8501…7a42.

Dashboard: https://agent-works-web.vercel.app/ · Integrate: see core's docs/INTEGRATIONS.md.