How to Audit a Multi-Agent CrewAI Crew: Find the Handoff That Broke the Output

Описание и применение

Руководство описывает, как добавить аудит к CrewAI с помощью ZizkaDB, создавая цепочку событий, позволяющую отследить каждый шаг, хэндшок и память агента.

Задача

Multi‑agent crews can produce wrong outputs because of subtle errors in handoffs that are not captured by standard observability tools.

Результат

The approach yields a tamper‑evident audit trail that links every task, handoff, and memory event, enabling causal debugging, compliance evidence, and time‑travel investigations.

Рабочий процесс

• Start a local ZizkaDB instance via Docker
• Install the ZizkaDB SDK and CrewAI logger
• Wrap the CrewAI kickoff and task loop with logger calls (log_kickoff, log_task, log_output)
• Log additional fields per task (role, sources, confidence, skipped actions)
• Record memory writes and supersedes relationships
• Use the dashboard for session replay and causal lineage queries
• Apply time‑travel debugging with db.at() to reconstruct past crew states

Доказательства

The logger records the crew lifecycle with three methods: log_kickoff for the start of the run, log_task for each task, and log_output for the final result.

Источник

The last line is the payoff. It walks back from the final output through every task to the kickoff, so a run becomes a causal tree and not a pile of unrelated lines.

Источник