Pydantic AI 2.20.0 Makes MCP Errors Recoverable, Not Fatal

Pydantic AI's 2.20.0 release stops a bare McpError from killing an entire agent run, per the GitHub release notes. That's the difference between one flaky MCP server aborting a production run and the agent routing around it and continuing.

The release notes don't say whether the agent retries the failing server or just routes past it. They also don't say how many dropped servers it takes before something upstream notices.

The same release changes how provider fallback works. Tool-search history now replays across provider boundaries instead of getting dropped. An agent that switches providers mid-run keeps its search context instead of starting over.

2.20.0 also adds Claude Opus 5 support. It starts capturing Anthropic's thinking tokens in usage details too, a fuller accounting of what a run actually costs.

A second release, 2.21.0, followed on July 30. It adds a narrower guardrail: a per-request input token limit that caps how much context a single call can burn.