Test agent orchestration with a deterministic mock model

Goal

Check workflow control logic with repeatable inputs and scripted model outputs.

Steps

1. Separate behavior owned by the application from behavior requiring a live model or remote service. [openai-testing]
2. Script model output with agents.testing.ScriptedModel([assistant_message(...)]); include function_call(...) for tool trajectories. [openai-testing]
3. Run workflow tests with the SDK’s test runner and mock model; disable trace export in test configuration. [openai-testing]
4. Assert call order, final output, and rejection/failure branches without relying on random model wording. [openai-testing]
5. Reserve provider- or network-specific behavior for isolated integration tests. [openai-testing]

How to verify

Repeated runs of the same script should produce the same tool sequence and result, with no live model request.

Troubleshooting

If a test unexpectedly calls a provider, verify the mock model is installed and tracing export is disabled.