Checkpointed DeFi Agents With On-Chain Accountability and Early Policy Enforcement

초록

Decentralized finance (DeFi) agents automate multi-transaction workflows such as swapping, lending, and vault management, but they also create process-level risk. A run can consist of individually valid calls while still becoming economically unsafe because an intermediate step leaves latent authority, weakens execution constraints, or accepts unverified external evidence. Existing defenses are often mismatched to this process-level risk. Off-chain preflight checks alone cannot protect against runtime deviations from the intended plan, and coarse on-chain allowlists are too weak to express the call-level intent that matters in DeFi. We present CheckpointAgent, a workflow-security architecture for checkpointed DeFi-agent execution. It composes manifest commitments, smart-account policy guards, post-state predicates, and attestation-gated advancement to constrain a run step by step and tie checkpoint advancement to verifiable evidence. Rather than judging safety only after a workflow finishes, CheckpointAgent checks whether each step remains consistent with the intended workflow and stops execution when the required conditions no longer hold. In the author-curated 27-scenario local-chain suite, the strongest evaluated setting preserves all 5 benign runs and prevents unsafe completion in all 22 adversarial runs, stopping them either through on-chain enforcement or through trusted-attestation advancement under the configured attester assumption. Under explicit trust assumptions and within the measured workflows and snapshots, checkpointed execution can materially reduce process-level risk without modifying target protocols.

키워드