Microsoft’s Azure AI Agent Server Invocations Python SDK has a new beta, version 1.3.0b1, with changes to voice handling, structured application data and request attribution. For developers building AI agent servers, this is a substantive set of new tools and behaviour, not just a version-number shuffle.
Microsoft AI Watch analysis
What happened
The release notes describe a beta dated 26 October. They add a standalone WebSocket route for voice conversations and align caller interruption with a named barge-in event contract. The notes also introduce structured-data types for Bridge protocol 1.0 and invocation-local original-subject assertions for HTTP, WebSocket and typed Voice handlers.
Our top picks
- A dedicated voice-conversation WebSocket route
Experimental voice servers can use a separate route alongside the existing invocations WebSocket endpoint. - A clearer caller-interruption contract
Voice handlers gain a defined barge-in event and callback, while playback snapshots and application-owned cancellation and history remain in the application’s hands. - Structured data for Bridge 1.0
New immutable types represent input data parts and completed structured-data output. - Original-subject assertions, with a deliberate boundary
Handlers can capture invocation-local assertions; passing model-attribution headers is an explicit per-call opt-in, not a change to model authentication. - More careful handling of malformed attribution
The notes describe classified errors for invalid metadata, including how the typed Voice WebSocket endpoints close the connection.
Why it matters
Voice is an awkward fit for interfaces designed around a single request and reply. A dedicated route and an explicit interruption contract give developers more defined pieces to work with, while keeping cancellation and conversation history under application control. Structured data and invocation-local context address a different, equally practical problem: making information available to handlers without quietly changing who owns it or how it is passed on.
The release also raises the minimum dependency to azure-ai-agentserver-core 2.3.0b2, which the notes say must be published first. This is a beta SDK update, so the changes are a preview for developers rather than evidence that a finished service has launched.
Our read
The best changes here are the ones that make responsibilities explicit: who handles an interruption, which route serves voice, and when attribution headers are passed to a model. That is less glamorous than promising an agent that does everything, and considerably more useful to someone building one. Developers should check the dependency requirement and treat the experimental voice and attribution features as beta components.
What to watch
- Whether the experimental voice route and barge-in contract settle before a stable release.
- How the new structured-data types are used in Bridge applications.
- Whether future releases expand the original-subject feature or keep model-header propagation opt-in.
Discussion spark: For AI voice applications, should caller interruption and conversation history be standardised by the SDK, or stay firmly under each application’s control?
Sources and evidence
- azure-ai-agentserver-invocations1.3.0b1 (27 October 2026, 12:02 UTC)
not affiliated with or endorsed by Microsoft