

Version 0.25.0 of OpenAI's Python Agents SDK removes the beta Runloop sandbox provider and changes checkpoint handling for Pydantic run contexts. The release notes advise preserving provider workspace data before migration and upgrading checkpoint readers before writers when multiple workers share stored state.
The GitHub release was published November 18 UTC; its changelog is dated November 17, consistent with the preceding U.S. Pacific evening. Those dates refer to the same release, not two separate launches.
A provider removal is also a data-migration question
The removal change deletes the Runloop integration. The release recommends moving to another supported provider, or remaining on 0.24.0 while migrating. It explicitly says provider-specific sessions and snapshots need a separate data migration.
Changing the package version or provider class therefore does not by itself transfer an existing workspace. That distinction is important for agent workloads whose useful state includes files or environment changes outside the conversation record.
Our analysis: durable state has several owners
An agent can have SDK-managed conversation state, application data and provider-managed workspace state at the same time. These may look like one continuing session from the interface, while having different storage and lifecycle rules underneath.
A migration needs to identify which layer owns each important piece of information. Otherwise, a workflow can resume its messages successfully while discovering that its files or external references are missing. The release note's separation of SDK and provider state is useful precisely because it prevents that assumption.
Compatibility between readers and writers matters for the same reason. When workers exchange persisted state, deploying a writer that emits a new representation before readers can interpret it creates a different problem from upgrading an isolated development environment.
Checkpoint fixes and configurable behavior
The release describes preservation of Pydantic defaults and in-place edits and support for resuming nested tool approvals with output guardrails under RunState schema 1.22. It also advises custom serialization for process-local resources. A database handle or similar live object should not be assumed to become portable merely because its surrounding context is checkpointed.
The SDK also changes a default compaction-session model and adds opt-in handoff recovery. Those are configuration and behavior changes, not proof that an existing application's recovery policy remains identical after upgrading.
The SDK source is MIT-licensed. Our repository profile links the implementation. Model APIs and sandbox providers retain separate access terms and costs. We inspected source and release notes without installing the SDK, migrating data, creating a sandbox or calling a model.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- openai/openai-agents-python ↗View project details
Original source and implementation review with labeled editorial analysis. No model inference, installation, paid service, API benchmark or independent security/robotics test performed.



