Realtime API with Multiple Tool Calls Results in No Voice Output

[sibblegp]

Describe the bug

When the SDK uses multiple took calls which return JSON (in this instance asking for film times which calls to search for the film and then for times of that film_id near a location), there is no voice output. When I say something back to it to say "What were those times again?" then it responds.

See this trace: Imgur

Note there is no audio out after the tool calls.

Debug information

• Agents SDK version: v0.2.1
• Python version: Python 3.13

Repro steps

Setup chained tool calls that should result in an audio out. Instead, no audio is generated.

Expected behavior

An audio out should be generated after multiple tool calls in sequence.

[Friedjof]

Same here

[norika1207-lab]

One practical way to make this easier to evaluate is to turn the report into a small contract test around the runtime boundary, not only the user-facing symptom.

For Realtime API with multiple tool calls results in no voice output, I would check three things: first, the failing input is preserved as a minimal fixture; second, the SDK emits the same observable event/result shape before and after the fix except for the intended change; third, the trace or serialized session contains enough metadata to explain what happened after a resume or retry.

The part that matters for agent systems is durability: if a tool call, memory write, handoff, or structured-output step fails, downstream code should not have to infer state from natural-language output. A focused regression test around that boundary would make the issue much easier to close with confidence.

[d87skg]

I've been working on a similar problem — capturing full execution traces with metadata for AI agents. I built an open-source tool that automatically records every tool call, LLM request, and error into a verifiable .evidence package:

pip install traccia-sdk
traccia intercept -- python your_agent.py
traccia verify traccia-.evidence
traccia diagnose traccia-.evidence

It works with the OpenAI Agents SDK and adds hash-chain verified metadata to every span. Might be relevant to what you're looking for.
https://github.com/d87skg/traccia

[IgorGanapolsky]

On realtime multi-tool runs that go silent on voice output, this usually is not a "model quality" problem — it is a control-plane race between tool fan-out and the audio sink.

What breaks most often in production:

1. Fan-out control — concurrent tool calls without a hard step/tool cap before the next audio turn is scheduled
2. Durable id + idempotency — the same tool result applied twice (or dropped once) when the session reconnects mid-turn
3. Observable long jobs — tool execution takes longer than the realtime turn window with no cooperative cancel / progress signal
4. Expected vs actual — the session marks "tools done" from an internal event while the external audio path never got a playable frame
5. Policy before action — no gate that refuses "continue speaking" when tool results are incomplete or partial

If you can share one sanitized failing trace (tool call order, expected audio event, actual last event id), the missing pillar is usually #3 or #4. Architecture-only note — not selling anything here.