Agent progress becomes durable workflow state

Microsoft has introduced an Azure Durable Task integration for Pydantic AI agents, allowing long-running work to survive worker restarts and deployment failures. The capability checkpoints model requests, function tools and MCP calls as Durable Task activities. When a run resumes, recorded results can be replayed instead of paying for the same model output and tool work again. Developers retain their Pydantic AI models, instructions and tools, adding durability as a capability rather than rewriting the application in a different agent framework.

Durability addresses a gap between demonstrations and production agents. A short chat turn can simply fail and be retried, but an agent that has researched documents, called several services and prepared a proposal may have accumulated minutes of work and material cost. Repeating every step wastes tokens and can produce a different result. Checkpointing creates a recoverable execution history so the orchestrator can continue from the last completed activity after transient infrastructure failure.

The integration uses Azure Durable Task Scheduler and the standalone Durable Task Python SDK. Model calls, ordinary tools and MCP tools become checkpointed activities by default when AzureDurableTaskDurability is attached to the agent. A Durable Task orchestrator schedules the bound capability, while a worker executes it. Developers can run the scheduler emulator locally in Docker before connecting to Azure, which supports a realistic development loop without requiring an Azure Functions application.

Checkpointing saves completed model and tool work

Recorded results are central to deterministic replay. The orchestrator must reconstruct control flow from history without re-running completed nondeterministic work. Model output is inherently variable, so keeping it in an activity result prevents replay from asking the model again. Application code inside the orchestrator still needs deterministic structure. Current time, random values and network calls should come through workflow-safe APIs or activities, otherwise a recovered execution may diverge from the history it is attempting to replay.

Retries can be configured around model and tool activities instead of implemented separately around each call. This is useful for rate limits and temporary network failures, but retry policy needs limits and error classification. Retrying an invalid request consumes time without changing the outcome. Backoff, maximum attempts and escalation should reflect the called service. Observability should show whether a successful workflow required repeated calls, because hidden retries affect both cost and reliability.

Human approval can wait without a live worker

Human approval is another strong fit. An agent can prepare a change, pause for hours or days and continue when a person responds without holding an activity or worker open. That separates compute time from calendar time and creates a natural approval boundary. The approval record should include exactly what was reviewed, with a digest or immutable version of the proposal. If the agent changes its plan after approval, the new action should require another decision rather than inheriting authority from an outdated preview.

Microsoft also supports coordinating several specialist agents within one durable workflow. An orchestrator can sequence tasks, run selected work in parallel and combine outputs. Durability does not decide how agents should collaborate or resolve disagreement. Teams need explicit contracts for inputs, outputs and failure. They should prevent an untrusted agent result from becoming instructions to another agent and record which model and tool version produced each checkpointed artefact.

Exactly-once side effects remain an application problem

The sharpest warning in Microsoft’s announcement concerns side effects. Durable recovery does not guarantee exactly-once execution for an external action. Work interrupted after creating an order but before recording the result may run again. Tools that change another system must be idempotent, usually through a stable operation key, conditional update or lookup-before-create pattern. Payments, notifications, infrastructure changes and ticket creation all require this protection. Checkpointing the agent is not a substitute for transaction design in the services it calls.

Azure Durable Task makes Pydantic AI agents more credible for processes that cannot finish in one request. It preserves completed work, supports deliberate waiting and gives operators a workflow history to inspect. Australian teams should test crash recovery at every activity boundary, including the uncomfortable moment after an external side effect. The integration is most valuable when combined with idempotent tools, bounded retries, versioned approvals and clear cancellation. Durability should mean the work resumes safely, not merely that it continues despite uncertainty.

Operational ownership must be explicit once an agent can outlive the web request that started it. Administrators need a way to inspect suspended runs, see the next eligible retry, cancel work and safely resume after a deployment. Retention rules should cover checkpointed prompts and tool outputs because durable state may contain personal or commercially sensitive data. Recovery is valuable only when the stored history is protected, observable and removable under policy. Teams should also test version upgrades against paused instances, because a changed schema or tool contract can make yesterday's checkpoint unsafe to replay tomorrow.