Source: github.com/pydantic/pydantic-ai/releases/tag/v2.62.0
The pydantic/pydantic-ai project released v2.62.0, which lets developers change thinking effort during a conversation through OpenAI's configuration_update and Anthropic's per-message output_config. These changes are recorded as ModelSettingsDeltaPart. The release also adds ModelRequestRejectedError, ModelAuthenticationError and ModelNotFoundError as model error categories.
Hosted MCP capabilities now prefix their tool names with the service name by default, and the prefix can be set with tool_prefix. The SummarizingCompaction summary is stored as a CompactionPart instead of a system prompt. CodeMode gains a monty_sandbox option, and monty_sandbox_url is deprecated.
Other fixes cover realtime sessions, AG-UI and Vercel AI message dumping, and OpenRouter and Gemini handling. The changelog also says GraphBuilder broadcast branches now start in .to() order rather than reversed, which can break replay of durable workflows paused across the upgrade.
Why it matters if you build agents
Agents that match on MCP tool names may see different names after upgrading, because the service prefix is now applied by default. Code that handles model failures can now tell rejected requests, authentication failures and missing models apart. Durable workflows paused before the upgrade should be checked for the GraphBuilder ordering change.
Get stories like this every weekday
One short email with the day's agent tooling news.