AI agents are increasingly acting on behalf of individuals or other agents to purchase services, exchange information, and delegate work. The rules an agent must follow are usually expressed as instructions, policies, or agreements, yet plain text cannot guarantee compliance because agents may err, be misled, or pursue their own interests. For an economy of such agents to be reliable, parties need enforceable agreements without having to inspect or trust each other's agents. We therefore introduce the Proof‑Carrying Contracts (PC2) framework, which makes agreed constraints enforceable throughout an agent’s interaction with a system. Before an operation is executed, the agent must provide a machine‑checkable proof that the agreed constraints hold in every case allowed by its model. The execution system, called Airlock, verifies the proof and decides whether to grant access. The parties specify the constraints, while the agent supplies the proof to justify its decisions and retains freedom to adapt its strategy, avoiding the need to design a separate permission policy and prove each decision safe. We prove that the constraints remain true at every instant during execution even for an adversarial agent, and we identify when this translates to safety in the real world. Reusable capability modules offer tools and guarantees that can be assembled into contracts. PC2 thus provides a foundation for an agent economy where agreements protect both delegators and executors.
Review: PC2 offers verifiable safety guarantees for AI agents, balancing flexibility with reliability, and constitutes a core technology for building trustworthy agent economies.