Auditability for agents: logging isn't only for debugging

When an agent autonomously calls tools, modifies data, or sends messages, a problem forces you to answer 'what did it do, on what basis, who authorized it'. Without a complete record, those questions can't be answered.

Auditability is not only for debugging but the basis of allocating responsibility and proving compliance — especially in regulated settings or with enterprise customers.

The pragmatic approach is to log an agent's inputs, decision basis, tools called, and results, and keep tamper-evident logs for sensitive actions.