Microsoft has released the first preview of azure-ai-agentserver-a2a, adding native support for A2A agents and persistent task storage in Azure’s Agent Server. The practical draw is that task records can survive a server restart, while the release notes make clear that this does not automatically recover task execution.
Microsoft AI Watch analysis
What happened
The Azure SDK release notes describe version 1.0.0b1 as the package’s initial preview. It adds an A2AAgentServerHost for hosting A2A v1.0 AgentCard and RequestHandler objects on Foundry container routes, alongside a FoundryTaskStore built on Agent Server Core’s generic state store.
The store supports caller isolation and native task listing, filtering and paging. Developers can also choose hosted Foundry storage or local in-memory defaults, and override the task store. A sample demonstrates retrieving a task after a server restart. The notes explicitly distinguish that persistence from automatic recovery of work that was running.
Other additions include request-scoped capture of call and user identifiers, transport-error classification, and OpenTelemetry spans for A2A operations. The public host and task-store APIs are marked experimental. Read the Azure SDK release notes.
Why it matters
A2A gives agents a way to communicate across systems; this preview adds Azure-specific hosting, storage and tracing pieces for teams building those services. Persistent task records can help an operator inspect work after a restart, while caller isolation and task-list controls matter when a service handles multiple users.
But persistence is not a magic restart button. The release notes say tasks can be retrieved after a restart, not that interrupted work resumes automatically. That distinction belongs in any deployment plan, along with the experimental status of the new APIs.
Our read
This is a meaningful developer preview because it puts hosting, task storage and observability into one Azure Agent Server package. The useful early test is whether teams can manage and inspect tasks safely in real deployments, without mistaking saved state for recovered execution. Preview means there is room to try it, and room for the edges to move.
What to watch
- Whether the A2A host and task-store APIs become stable, and what changes on the way.
- How hosted storage and caller isolation behave in deployed services.
- Whether later releases add a supported route for recovering interrupted task execution.
Discussion spark: For agent services, is durable task storage enough for a useful first release, or should automatic recovery be part of the minimum bar?
Sources and evidence
- azure-ai-agentserver-core2.3.1b1 (5 November 2026, 09:11 UTC)
- azure-ai-agentserver-a2a1.0.0b1 (5 November 2026, 09:11 UTC)
not affiliated with or endorsed by Microsoft