Evidence / Agent audit trail
If an agent acts,
keep the why.
An AI agent audit trail is more than a log of tool calls. It is the smallest complete record of what was proposed, what evidence was available, what authority applied and why the system allowed, deferred, escalated or blocked the action.
01 / Minimum record
The trace should
answer five questions.
- What was proposed?Capture the action, target, scope and the agent or service that proposed it.
- What did the system know?Preserve relevant evidence, timestamps, provenance and material uncertainty.
- What authority applied?Record the mandate, permission boundary, policy version and approval state.
- What was decided?Make the outcome explicit: allow, defer, escalate or block.
- What happened next?Link the execution result, human intervention, correction or rollback.
02 / Common gap
Observability is
not governance.
Tracing prompts, tokens and tool calls helps engineers debug a system. It does not, by itself, prove that the action was authorised or that its evidence was sufficient. Governance adds a decision boundary and a reason that can be reviewed independently of the model's confidence.
That distinction matters most when actions touch money, production systems, customer records, access rights or irreversible state.
03 / A useful shape
Preserve context
with the outcome.
trace_id = trace_01J...
proposal = { action, target, scope }
evidence = { source, observed_at, confidence }
authority = { actor, mandate, policy_version }
consequence = { reversibility, severity, uncertainty }
evaluation = { capability, evidence, authority, consequence }
decision = { outcome, reason, escalated_to }
execution = { status, result, completed_at }Practical standard: a new operator should be able to reconstruct the decision without replaying the entire agent conversation.
04 / Continue