Agents
Pydantic AI 2.20.0 makes a bare McpError recoverable instead of fatal and replays tool-search history across provider boundaries
The July 29 release stops a bare McpError from an MCP server killing the run, which is the difference between one flaky MCP server aborting an agent and the agent routing around it. It also replays tool-search history across provider boundaries instead of dropping it — relevant to anyone doing multi-provider fallback — adds Claude Opus 5 support and OpenAI Responses API reasoning.context (defaulting to all_turns) for the gpt-5.4/5.5/5.6 families, and captures Anthropic thinking tokens in usage details. Version 2.21.0 followed on July 30 adding per_request_input_tokens_limit to UsageLimits.
↳ Follow the thread