OpenAI Realtime Agents
Working baseline established for the selected voice, tools and agent-handoff workflow.
A spoken request, an agent handoff, and a recorded response.
The native application exchanged controlled voice over WebRTC, handed a spoken request to a haiku agent, and routed an example account inquiry through a supervisor and the repository’s account tool. Factory also observed response interruption and session closure.
Voice sessions verified
Expectations met
Evidence records
What was verified
All five declared behavior expectations were met in the selected workflow.
- Controlled input and output audio crossed the native WebRTC connection and matched the fixed reference recordings.
- A spoken request transferred control from the greeting agent to the haiku agent and displayed the declared response.
- The supervisor followed the native account-tool call and returned the repository’s example balance of $42.17.
- The selected interruption and disconnect workflow cleared the active session and released its audio resources.
- The selected workflow met its declared request, media and execution budgets.
Coverage of this run
Two controlled voice sessions covered sequential handoff and supervisor delegation, followed by response interruption and closure.
- Native voice and event transport: Established with fixed audio and a Factory synthetic provider
- Sequential agent handoff: Greeting agent to haiku agent observed
- Supervisor delegation and account tool: Established with fictional input and repository example data
- Response interruption and session closure: Established for the selected workflow
- Application verification: 49 of 49 stated criteria passed
- Independent consequence verification: 50 of 50 stated criteria passed
- Behavior verification: 14 of 14 verifier criteria passed; 5 of 5 workflow expectations met
- Live-provider speech, recognition and answer quality: Further verification scope
- Production networks, persistent account data and customer authorization: Further verification scope
Each count belongs to its stated verification scope. The account inquiry uses fictional input and the repository’s fixed demonstration data.
The recorded application includes one disclosed recorder-cleanup repair. The repository’s agent definitions and API handlers retain their original behavior.
Supporting evidence
Five reviewed browser captures from the recorded run, followed by curated summaries of its seven selected evidence records.
Accepted workflow result
The selected run passed F0 through F3. Two voice sessions included two spoken submissions and one text submission. Automatic greetings and native continuations were counted separately.
SDK verification
All 42 stated SDK criteria passed. The installed SDK, agent definitions, native audio and event transport, tool joins, and closure were checked with the controlled provider.
Application verification
All 49 stated application criteria passed, covering the production build, readiness, native interface, controlled voice, moderation, selected tool activity, budgets and five screenshots.
Consequence verification
All 50 stated consequence criteria passed. An independent review reopened the captured application evidence and connected handoff, the example account tool, recorded audio, interruption and closure.
Behavior verification
Five frozen workflow expectations were met across media, handoff, supervisor tools, interruption and budgets. Fourteen verifier criteria confirmed their selection, evidence binding and repeatable assessment.
Visual review
Five recorded screens show connection, handoff, the supervisor result and the closed session. The separate media records establish audio delivery and content identity.
Applied engineering checks
Five selected checks were applied and passed during preparation, capture, native-protocol verification, evidence publication and recorded-workflow review.
Captured voice
Fixed speech from the Factory synthetic provider, captured after native WebRTC delivery. These recordings demonstrate controlled media delivery. The sample account details are fictional.
Agent handoff
Controlled greeting and haiku response captured from the native output track.
Supervisor response
Controlled greeting and example account response captured from the native output track.
Recommended next verification
Verify the configured live-provider workflow.
Select the live provider and model, then measure speech recognition, response quality and voice latency against explicit criteria.
Extend voice lifecycle coverage.
Add device changes, delayed microphone permissions, reconnects and varied interruption timing to the selected workflow.
Verify production account behavior.
Use an authorized production-equivalent environment to evaluate network conditions, persistent data and customer access rules.