Source: github.com/pydantic/pydantic-ai/releases/tag/v2.60.0
The release lists several compatibility changes. RetryPromptPart is deprecated in favor of ToolReturnPart with outcome set to retried, alongside a new RetryFeedbackPart. The MCP capability now derives its default id from the server URL's hostname, and two servers on the same host are rejected unless each has an explicit id.
New features include DockerSandbox, which runs a workspace in a local Docker or Podman container, and BoxLiteSandbox, a local BoxLite micro-VM workspace. RelevanceCompaction keeps, reduces or drops each older history step based on a judge model's ranking. MCPToolset can now take auth and headers as functions of the run context, so each run connects as its own user.
Other changes cover Bedrock, OpenAI, Anthropic and realtime voice sessions, along with many bug fixes. Bedrock's default maxTokens now matches Claude's maximum output, and Temporal durability adds continue-as-new support.
Why it matters if you build agents
Code that builds RetryPromptPart should move to ToolReturnPart(outcome='retried') or RetryFeedbackPart, since the old type is now deprecated.
If you run several MCP servers on one host, give each an explicit id, because the release now rejects duplicates derived from the hostname.
The run-context hooks for MCPToolset let a multi-user agent authenticate each user separately instead of sharing one connection.
Get stories like this every weekday
One short email with the day's agent tooling news.