Agent action tracing

What it does

It makes agent behaviour examinable after the fact. It is what allows an incident to be investigated, a decision to be explained to the person it affected, a rogue pattern to be detected, and the other controls in this category to be evidenced. Engineering operates it continuously; the agent's owner is accountable for the record.

How to implement it

• Capture for each task: the request and requester, the agent's plan, each tool call with parameters and result, any approval, and the final output.
• Carry one identifier across the task, including delegation to sub-agents.
• Record the model, prompt and component versions in force.
• Store traces in the organisation's own systems, protected from alteration.
• Set retention by the period in which the action may be disputed.
• Feed traces to monitoring that looks for departures from mandate.

Evidence it operates

Design: a tracing standard listing the required fields and retention. Operation: for a sample of past agent actions, the full sequence from request to effect can be reconstructed from stored records alone; trace volumes reconcile to action volumes; an attempt to alter a stored trace is detected.

Where it fits

Built in at design and running in production. It extends AI system monitoring and logging to the reasoning and delegation that ordinary logs omit, and supplies the audit trail for approval gates.

Limits

A trace records what the agent reported about its reasoning, which may not be how the output was actually produced. Traces contain whatever the agent handled, including personal and confidential data, and need the same protection and retention discipline as the data itself. Tracing detects; it stops nothing by itself.