An agent with shell access can often delete or edit its own local history, and recent research shows most harnesses do not prevent it. Capture traces through an interception layer the agent cannot reach, and anchor them off host.
Read in this order
- PaperLLM Agents Can Easily Tamper With Their Own TracesStart here: the evidence that agents can rewrite their own traces.
- PaperAuditable AgentsA design for agents whose actions can be audited afterwards.
- OfficialArticle 12: Record-keeping (EU AI Act Service Desk)The legal record keeping duty for high risk AI systems.