The report
A JFrog research proposal called Agentic Process Evidence addresses what a Git commit cannot ordinarily show. A source change identifies a resulting artifact, but not the exact agent session, model, tools, external data and approvals that shaped it. The proposal ties process records to artifacts through signed attestations and suggests independent delivery gates that evaluate those records against policy.
The difference between recording and enforcing matters. A signed statement can show that a named collector attested to a specific set of bytes. It cannot prove that an agent's testimony is complete, that every shell call was captured or that an omitted tool action never occurred. Similarly, a gate must run outside the agent's authority if it is to prevent the agent from approving its own output. The paper supplies architecture and illustrative use cases, not independently demonstrated elimination of production incidents.
The canonical design dossier specifies the remaining questions around compromised runtimes, log completeness and the privacy of captured context. A useful implementation trial would compare signed session logs to host-level effects and deliberately remove one event to test whether the release gate fails closed.