An agent can pass a final-output check while taking harmful or unauthorized steps along the way. Multi-step testing probes intermediate actions and interactions, then measures how errors propagate through the workflow. section 6.7.1
A practical testing and monitoring pattern:
- Probe different steps with wrong or unexpected inputs; test agent interactions and the whole system when agents collaborate. section 6.7.1
- Measure error propagation, and scale testing and online monitoring to the risks of the agent. section 6.7.1
- Add programmatic runtime checks that halt the workflow when an unauthorized tool call is attempted. section 6.7.1
- For long-running agents, track intent across steps rather than judging each action in isolation. section 6.7.1
