Microsoft’s Azure AI Projects Python SDK 2.9.0 adds experimental tracing for realtime voice agents and lets developers choose WebSocket or WebRTC connections. The release also changes some preview APIs, so projects using those interfaces may need adjustments.
Microsoft AI Watch analysis
What happened
The 2.9.0 release notes, dated 16 October, describe several changes. Voice Agents can now be traced through AIProjectInstrumentor, including session, connection and response spans, tool and lifecycle events, audio-size totals and existing generative AI response metrics. Message content is not captured by default, and existing tracing and content-recording settings still apply.
Developers can also select WebSocket or WebRTC when connecting to realtime voice agents. The SDK builds the transport query internally. Other additions include preview support for Microsoft Copilot managed harnesses and toolsets, plus fixes to cancellation handling in long-running Azure AI Projects operations.
The breaking changes are confined to preview APIs. They include renamed Teams Phone Extensibility classes and enum values, removal of two response base classes, and changes to persisted response fields and metadata.
Why it matters
This gives developers more visibility into realtime voice-agent sessions and another connection option, while the default tracing behaviour avoids recording message content. That is useful operational detail for teams building and diagnosing voice applications, not merely a new label in a version number.
The preview API changes deserve a look before upgrading: code that relies on the affected names or response structures may need updating. Microsoft says the breaking changes affect preview APIs, rather than presenting them as changes to stable interfaces.
Our read
The strongest addition is the combination of voice-session tracing and transport choice. Better traces can make a realtime system easier to understand when something goes wrong; WebRTC gives developers another route to connect. Do check the preview API migration points before updating, though. “It’s only preview” is not quite the same as “nothing in your build depends on it”.
What to watch
- Whether the experimental tracing and transport options move beyond their current status.
- Which preview API names and response fields need changes in projects upgrading to 2.9.0.
- Whether future releases add further voice-agent diagnostics or connection options. Activity teaser Azure AI Projects 2.9.0 adds experimental tracing for realtime voice agents and lets developers choose WebSocket or WebRTC connections. Message content is not captured by default, though existing tracing settings still apply. The release also changes some preview APIs, so upgrading may mean code edits for teams using them. The full story covers what changed and what developers should check. Is better visibility into voice-agent sessions worth adopting an experimental feature, or would you wait for it to mature?
Discussion spark: Would you adopt experimental voice-agent tracing now for the extra visibility, or wait until the feature is no longer experimental?
Sources and evidence
- azure-ai-projects2.9.0 (17 October 2026, 03:44 UTC)
not affiliated with or endorsed by Microsoft