Observability
How tracing and evals help assess Deep Agent investigations.
What this is
Observability records how scanner, model, and citation checks behave during validation.
When to use it
Use it when comparing model behavior, debugging investigation quality, or running live validation.
Command or example
uv run --extra deep-agent --extra phoenix agent-permit live-validate . --phoenixOutput to expect
Phoenix traces and local metrics can show model calls, token usage, cache hit ratio, and citation outcomes.
How to interpret it
Observability is for quality and cost review. It does not change deterministic permit status.
Common mistakes
- Treating tracing as product value by itself.
- Uploading sensitive traces without policy.
- Ignoring citation pass/fail because the prose looks convincing.