Untraceable agent actions

What it is

To account for an agent's action the organisation needs to know what was asked, by whom, what the agent planned, which tools it called with which parameters, what came back, and under whose authority it acted. Ordinary application logs capture the final call at most. This entry is the agent-specific case of record-keeping failure: the reasoning and the delegation chain are part of the record, and they are lost unless deliberately captured.

How it happens

• Only the final output is stored; intermediate steps are discarded.
• Agent actions are logged under a shared account.
• Traces are kept by the model provider and expire.
• Sub-agents log separately with no common identifier.

Why it matters

After an incident the questions are what happened, why, and who is responsible. If they cannot be answered, the organisation cannot correct the cause, cannot tell affected people or a regulator what occurred, and cannot show that its controls worked.

Warning signs

• A past agent action cannot be replayed from records.
• Logs show the action and omit the instruction that led to it.
• Retention of traces is shorter than the period in which disputes arise.
• Different agents' logs cannot be joined.

Who owns it

The agent's owner is accountable for its record. Engineering implements tracing. Records management sets retention. Internal audit tests that a sample of actions can be reconstructed.

In practice

A customer disputes a change of address made on their account. The system shows that an agent made the change. It does not show which conversation led to it, what the agent was told, or whether verification took place. The same pattern generalises to any agent whose actions may later need to be explained.