AI Agent News/stories/openai-agents-python-sdk-v0-25-0-removes-the-runloop-sandbox-provider.md

Frameworks

OpenAI Agents Python SDK v0.25.0 removes the Runloop sandbox provider

OpenAI's openai-agents-python SDK v0.25.0 removes the beta Runloop sandbox provider and changes how Pydantic run contexts are checkpointed. Developers should move Runloop workloads to another provider, such as Modal, and upgrade resume workers before enabling new checkpoint writers.

Source: github.com/openai/openai-agents-python/releases/tag/v0.25.0

OpenAI's openai-agents-python SDK released version 0.25.0 on 18 November. The release removes the beta Runloop sandbox provider, including its extra and exports. Users of that integration need to preserve their workspace data before upgrading, and the release notes name Modal as a supported replacement. Existing Runloop session IDs and snapshots cannot be resumed directly in Modal.

Checkpoint handling changes as well. Run state moves to schema 1.22 to track approvals owned by nested agent tools, and Pydantic run contexts now keep default values and in-place edits. Workers running v0.24.0 reject 1.22 checkpoints, so resume workers should be upgraded before new writers are enabled.

Other changes include a new default model for OpenAIResponsesCompactionSession, which moves from gpt-4.1 to gpt-6-luna. Runner handoffs gain an opt-in failure_error_function for recovery, and MCP v2 Streamable HTTP now allows concurrent requests. The demo REPL can also approve or reject individual tool calls.

Why it matters if you build agents

Teams on the Runloop integration must migrate providers or stay on 0.24.0 until they do. Anyone who relies on the compaction session's default model should set model to gpt-4.1 to keep the previous behavior. Retries for MCP Streamable HTTP tool calls should only replay calls that are safe to repeat.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe