Pydantic AI Updates: October 23, 2026

1. Pydantic AI 2.60.0 Adds Docker and BoxLite Sandboxes, RelevanceCompaction, and SmartSearch

Pydantic AI. Pydantic AI v2.60.0 adds a DockerSandbox harness capability that runs the agent workspace in a local Docker or Podman container, a BoxLiteSandbox backed by a local BoxLite micro-VM, and the option for SmolSandboxBackend to reuse a VM the application already has. New capabilities include RelevanceCompaction, which keeps, reduces or drops older history steps based on a judge model’s ranking, SmartSearch for plain-English code search with a pluggable judge model, and FileUnderstanding. MCPToolset can now take auth and headers as functions of the run context so each run connects as its own user, and TemporalDurability gains support for Temporal continue-as-new. Source

2. v2.60.0 Ships a Long List of Compatibility Changes and Security Hardening

Pydantic AI. The release deprecates RetryPromptPart in favor of ToolReturnPart(outcome='retried') and RetryFeedbackPart, deprecates OpenAIDecisionsProvider, adds a 'max' unified thinking effort level, and defaults BedrockConverseModel’s maxTokens to Claude’s maximum output, streaming such requests behind the scenes. Hardening changes include stripping request-level instructions from untrusted history in sanitize_messages and server-managed UI adapters, keeping BrowserUse secrets on allowlisted sites and out of tool results, matching FileSystem denied patterns case-insensitively while protecting .git files, and routing framework tags through a wrapper so a </system> inside an instruction or tool name cannot end its block. The MCP capability’s default id is now derived from the server hostname, and two servers on one host without explicit ids are rejected. Source