OpenAI Agents SDK for Python 0.19.3 lands 36 fixes concentrated on guardrail reporting, session durability, and tool-call ordering
Released 2026-08-04, v0.19.3 is almost entirely correctness work on the failure paths agents hit in production: output guardrail results are now reported when a tripwire aborts a run, tool guardrail results are reported for streamed runs, committed tool session records survive a streamed output guardrail trip, and a failed SQLiteSession insert now rolls back instead of leaving partial state. Ordering bugs get a cluster of their own, with fixes preserving tool call and output order when deduplicating inputs, keeping input item order when collapsing duplicates, and stopping handoff calls from being emitted as streamed `tool_called` events. Sandbox handling was hardened (default snapshot path resolution, apply_patch move source), and MCP tool and prompt listings now auto-paginate rather than silently truncating.
↳ Follow the thread