A code diff cannot explain the whole process
The Agentic Process Evidence paper argues that conventional software records leave gaps when an AI agent selects dependencies, calls tools or changes infrastructure. A commit can show the resulting edit without preserving which model, harness or tools produced it, or what approval applied.
The authors propose Agentic Process Evidence, or APE: a signed, tamper-evident, machine-readable attestation connecting an agentic session with its outcome. Its focus is process-aware logging rather than replacing the artifact record already kept by version control.
Connect provenance with policy checks
The paper describes support for an in-toto-style attestation framework and the use of evidence in policy-as-code controls. It frames supply-chain risk, policy guardrails and task alignment as related procedural problems.
A software bill of materials describes components in an artifact. A process record asks another set of questions: which session introduced a change, which tools participated and whether a person reviewed the result. The two records can therefore be complementary rather than interchangeable.
A signed statement still needs a trustworthy source
Tamper evidence helps establish whether a recorded statement has changed. It does not independently establish that the statement is complete, that the agent acted correctly or that the recording system observed every relevant event. Nor does the proposal itself certify compliance with a particular legal requirement.
Consider reviewing a dependency change made by an agent. The useful evidence would connect the resulting commit to the actual session, its tool activity and the relevant approval, instead of relying only on a narrative summary written afterward. APE presents an approach to retaining those connections. Its value depends on the quality of event collection, the verification process and the policy that uses the record, not merely on attaching a signature to an output.
Comments