Tool exceptions and serialization guardrails
OpenAI released version 0.24.0 of its Python Agents SDK on October 15, 2026. The minor release introduces a run-wide tool exception formatting mechanism for decorated functions. Custom error handlers can now receive a tool_exception kind along with an optional error object, preventing broken tools from defaulting to missing-tool handling routines. Per-tool configurations and MCP handlers retain their existing precedence.
The update also advances durable state serialization to schema 1.21. The updated reader rejects historical checkpoint graphs that contain same-name non-root agents whose identity depended on hook instance attributes. Teams managing durable state must route ongoing legacy executions through their original SDK and agent configurations, while assigning unique names to agents in new workflows.
Session persistence and environment isolation
The update refines recovery boundaries during multi-tool execution. When one tool fails, the framework preserves successful sibling tool outputs rather than discarding the entire sequence. Automatic memory compaction now pauses while human approvals remain pending, and remote conversation session operations are batched to stay within service request limits.
Local security controls receive adjustments under macOS profiles, where the shared sandbox environment now denies Launch Services requests to prevent local commands from spawning host applications. The release notes clarify that local sandboxing does not provide full host or network isolation, directing developers toward hosted or containerized environments when stricter isolation is needed.
Our assessment
Teams managing long-running agent state should verify that non-root agents have unique names before creating new durable sessions. Existing custom error formatters must be audited to ensure that the new tool_exception classification returns safe messages rather than leaking raw exception details directly to the model.
Source: OpenAI original publication.
Read the original material
OpenAI (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
