Back to catalog / Agent workflows

Evaluate full-duplex voice-agent task outcomes

Evaluate continuous voice interaction, actual tool calls, backend state, and the final answer together in an end-to-end task report.

Goal

Evaluate continuous voice interaction, actual tool calls, backend state, and the final answer together in an end-to-end task report.

Use CRAWL synthetic audio for controlled single-turn requests. [src-openai-evaluate-full-duplex-voice-agent-task-outcomes]
Use WALK recordings to test audio robustness and interaction behavior. [src-openai-evaluate-full-duplex-voice-agent-task-outcomes]
Use RUN simulated multi-turn callers and verify backend task state against the spoken result. [src-openai-evaluate-full-duplex-voice-agent-task-outcomes]

For each case compare caller intent, tool trace, final backend state, and spoken response; report failures separately by mode.

If speech sounds correct but backend state is wrong, inspect executed tools and state assertions; for replay drift, verify audio-to-case mapping.

Voice agent under test, synthetic audio, isolated test backend, and credentials for the selected generation/replay APIs.

Evaluation sends audio and may execute test-backend actions; do not connect live write systems or upload unapproved recordings.

Transcript quality or naturalness cannot replace validation of application state and task completion.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT