No Certificate, No Execution: Certified Traces as a Foundation for Trustworthy AI Agents

Trustworthy AI agents should require formal certification before taking action, not just rely on better models or safety filters after the fact. The paper argues that AI systems in high-stakes domains (healthcare, finance, policy) should only execute actions that have been formally verified as permissible under relevant policies—similar to how a building permit must be approved before construction begins.

Individual actions can be safe while their combination is not. The paper emphasizes that even if each step an AI agent takes seems reasonable in isolation, the sequence of actions together could violate policies or cause harm. For example, a series of individually appropriate medical recommendations might contradict each other or violate treatment protocols when combined.

Visible reasoning alone is insufficient for trustworthiness. While showing an AI's "chain of thought" helps humans understand its reasoning, the paper argues this is not the same as proving the AI followed proper procedures. A system could show logical reasoning but still violate required policies, credentials, or approval processes—making formal certification necessary alongside transparency.

Execution traces should be complete, structured records submitted for approval. Before an AI agent acts, it should create a detailed record including intended steps, evidence, tool calls, required approvals, and execution conditions. This structured approach enables systematic verification rather than relying on post-hoc audits after actions have already occurred.

Permission and capability are different. Just because an AI system can generate or compute an action doesn't mean it should be allowed to execute it. The paper argues for a clear separation between what AI agents can propose and what policies actually permit them to do.