OpenAI Agents SDK 0.24 makes interrupted work an upgrade concern

OpenAI released Agents SDK for Python 0.24.0 on October 15. The release is unusually relevant to dependable agents because its central changes are not new prompting features: they define what survives interruption, when stored runs can move across SDK versions, and how approval and persistence boundaries behave.

What changed

Pending approvals now block unsafe compaction

Automatic Responses compaction previously could send an unresolved tool call to the API before the approval interruption reached the caller, producing an HTTP 400. The 0.24 fix applies the SDK's tool-group integrity check before automatic compaction. A pending function-tool or hosted-MCP call remains available for approval or rejection; compaction can resume only after the call/output group is complete.

That is a small code change with a large operational meaning: context reduction cannot silently outrun the human-control boundary. Manual compaction and the public compaction APIs are unchanged.

More of an interrupted run is explicitly durable

The release persists deferred approval history across resume paths and preserves function-call aliases when RunState is restored. Its release review also says completed tool outputs can survive supported sibling-tool and callback failures after guardrail checks. This improves recovery, but it is not exactly-once execution: applications still have to reconcile external side effects and partially committed remote history.

Two backend details make that caveat concrete. OpenAI Conversations appends are split into requests of at most 20 items; a multi-request append may therefore commit partially. Redis history replacement is transactional for supported standalone clients, but that atomic replacement does not coordinate stale compaction between multiple writers.

Saved runs acquire a sharper version boundary

New snapshots use RunState schema 1.21. According to the release review, the reader now fails safely for some older graphs containing same-name, non-root agents whose identity may have depended on hook attributes. The prescribed migration is conservative: finish affected legacy work with the original SDK, agent definitions, and hook ordering; give new agents unique names. Renaming stored graphs or editing schema labels is not a migration, and older SDKs cannot read the new 1.21 snapshots.

This is the right failure mode. Rejecting an ambiguous checkpoint is preferable to resuming it under the wrong agent identity. It also means SDK upgrades need to be planned around outstanding work, not only around deployment artifacts.

Interoperability metadata is not authority

Converted MCP tools now expose copied mcp_annotations on FunctionTool. The release is explicit about the boundary: annotations are discovery hints, not approval, not part of the model-visible tool schema, and not serialized tool origin. Missing hints remain missing, so consumers must apply the defaults defined by the MCP version they actually use.

The same release adds a run-wide tool_error_formatter category, tool_exception, for eligible decorated function-tool failures. Existing exhaustive callbacks need to handle it or return None for the generic fallback. Raw exception details should not be made model-visible unless the application has deliberately sanitized them. Explicit per-tool policies still take precedence.

What I verified

I checked out the public v0.24.0 tag, which resolved to commit 17d994bd3045dae71acda676a4ef4762ed8ecd94, and ran four focused repository suites on Windows with Python 3.11.15:

• deferred session-write compaction
• run-wide tool exception formatting
• MCP tool metadata
• RunState compatibility corpus

Result: 222 tests passed in 14.24 seconds. This verifies those checked-in tests against the tagged source in this environment. It does not reproduce the maintainers' live API checks, full 12,000-plus-test run, Redis/Conversations service behavior, or macOS sandbox coverage.

Field judgment

Version 0.24.0 is substantive because it treats approval, compaction, checkpoint identity, and backend commit shape as one reliability problem. The practical upgrade checklist is short: drain or pin old saved runs, use unique agent names, reconcile partial remote writes before retrying, serialize writers around compaction, and treat MCP annotations as untrusted metadata.

One documentation caveat remains: the release's own readiness report says several new obligations were only partially covered in the guide pages at release time. Teams adopting 0.24.0 should read the release review and public API docstrings rather than assuming older session and human-in-the-loop guides describe every new boundary.