Source: github.com/openai/openai-agents-python/releases/tag/v0.24.0
RunConfig.tool_error_formatter can now receive a new kind, tool_exception, along with an optional error field. Explicit per-tool handlers keep precedence over the run-wide callback, and the default message stays unchanged when no custom formatter is set.
New RunState snapshots are written with schema 1.21, which older SDKs cannot read. The reader also rejects some older graphs containing same-name non-root agents, and the changelog advises finishing those tasks on their original SDK version and agent code, then using unique agent names for new durable runs.
Recovery now keeps more completed work, including deferred approval history and tool outputs when a sibling tool fails. Automatic compaction waits for pending approvals, and OpenAI Conversations session writes are split into batches of at most 20 items. Converted MCP tools also expose copied mcp_annotations metadata.
Why it matters if you build agents
Developers who customize tool_error_formatter need to handle the tool_exception kind, because a formatter that treats every non-rejection as a missing tool can now return misleading text for a crashed function.
Teams with durable runs should finish or isolate older saved tasks before upgrading, since pre-1.21 SDKs cannot load new snapshots. MCP annotations are untrusted discovery hints and do not grant approval.
Get stories like this every weekday
One short email with the day's agent tooling news.