Pydantic AI v2.57.0 (2026-10-15)
Pydantic AI v2.57.0 introduces MetaProvider for Meta Muse models, an automated ApprovalReviewer for tool calls, MCP Apps integration forwarding to AG-UI and Vercel AI frontends, and several compatibility and session core adjustments.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
Pydantic AI v2.57.0 introduces multiple features and bug fixes across real-time, MCP, and model providers. Key compatibility notes include attaching namespaced origin and tool call id metadata to BackgroundTools completions, migrating GPT-Live and Gemini Live realtime history, usage, and reply waits to the session core, and honoring disable-model-invocation and user-invocable flags in Skills. Notable features include MetaProvider support for Meta Muse models using the Responses API, an ApprovalReviewer harness allowing models to evaluate low-risk tool calls prior to human review, MCP Apps tool call relaying and resource reading via MCPToolset, and frontend forwarding to AG-UI and Vercel AI. Fixes address token compaction history retention, prompt caching behavior, DBOS durable execution steps, and provider-specific quirks across Gemini, OpenAI, and xAI.
可执行摘要
Pydantic AI v2.57.0 adds MetaProvider for Meta Muse via Responses API, MCPToolset enhancements for MCP Apps, and ApprovalReviewer harness. Note compatibility changes involving namespaced origin/tool call IDs in BackgroundTools, and session core routing for GPT-Live and Gemini Live realtime history, usage, and reply waits.
- Agent 实用度
- 77/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Verify tool call handlers if using BackgroundTools, ensuring compatibility with namespaced origin and tool call id metadata.
- Update realtime implementations using GPT-Live or Gemini Live to account for history, usage, and reply waits being retrieved from session core.
- Review custom Skills definitions to ensure proper handling of disable-model-invocation and user-invocable configurations.