Pydantic AI v2.60.0

Summary

Pydantic AI 2.60.0 (2026-10-22, from the release feed) adds two local sandboxes for its agent harness (Docker/Podman and BoxLite micro-VMs). It also adds two harness abilities that use a judge model: plain-English code search and relevance-based history compaction. MCP toolsets can now authenticate each run as its own user.

Key points

• Sandboxes: DockerSandbox runs the agent's workspace in a local Docker or Podman container. BoxLiteSandbox uses a local BoxLite micro-VM. SmolSandboxBackend can reuse a VM the application already has.
• SmartSearch: a harness capability for code search in plain English, with a pluggable judge model.
• RelevanceCompaction: a judge model ranks each older history step, which is then kept, shortened, or dropped. This is an alternative to plain summarising.
• FileUnderstanding: a new capability for handling files.
• MCPToolset: auth and headers can be functions of the run context, so each run can connect to an MCP server as a different user.
• Other: continue-as-new in TemporalDurability. A lockdown option on the GitHub capability withholds public content from users without push access. Also new: Bedrock profiles for Grok 4.6/4.7, a configurable MoonshotAIProvider base URL, and native Responses input items for OpenAIDecisionsModel.

Why it matters

Local container and micro-VM sandboxes make it practical to run a Pydantic AI coding agent on your own machine without exposing the host. Per-run MCP auth is needed for multi-user apps that call MCP servers on a user's behalf.