What should an audit log capture when an AI agent changes Xperience?

Xperience already has event logging, content versioning, workflows, and source-controlled CI/CD for different kinds of change history. Agent-driven changes add another layer, though: the person requesting a change, the model or agent planning it, the MCP or API operation executing it, and the human approving it may all be different actors.

For content or configuration changes made by an agent, what is the minimum audit record a team should retain? I would expect the environment, initiating user, agent client and version, approval identity, intent or prompt summary, exact tool or API operation, affected object GUIDs, before-and-after diff, timestamp, correlation ID, result, and a rollback or content-version reference. Storing a short intent summary instead of the full prompt may also reduce the chance of retaining secrets or unrelated context.

Should this record live inside Xperience, an external observability platform, CI/CD, or all three? For destructive or production-impacting operations, should an immutable audit entry be written before the change is committed so the approval, execution, and resulting Xperience version can be traced end to end?

Tags:
AI Security KentiCopilot MCP

Answers

Mike, your field list looks thorough to me. My thoughts on the two closing questions:

For where it lives, I'd lean toward keeping the main record outside Xperience, in something like an external observability tool, so the agent or credentials being audited can't touch it. The Xperience event log and content versions can just be referenced by ID. For a smaller team, though, the built-in logging plus CI/CD history is probably a fine place to start.

For write-before-commit, I think it makes sense for destructive or production-impacting operations, then updating the entry with the result and version afterward so an interrupted operation still leaves a trace. For low-risk content edits it's probably more effort than it's worth.

Also curious whether anyone records which skills or instructions the agent was given. That seems like it would help explain odd behavior later.

To response this discussion, you have to login first.