How do I troubleshoot an agent that fails intermittently when using a framework-agnostic runner?

Start by reviewing the verifiable execution traces from the runner to isolate where the failure occurred. Check if the step was replayed after a crash, and whether external calls were idempotent. If the agent framework itself has a bug, the runner's traces help you pinpoint the issue without affecting other runs. Ensure tool authorization policies aren't blocking legitimate calls.