Source: github.com/pydantic/pydantic-ai/releases/tag/v2.57.0
Pydantic AI v2.57.0 adds a MetaProvider for Meta Muse models, selected with the meta: prefix and using the Responses API. A new Harness ApprovalReviewer allows a model to approve low-risk tool calls before a person sees them. Requests are now priced at their reported service tier, and EvaluationReport gains a dataset_name field.
On the MCP side, tool results from MCP Apps are forwarded to AG-UI and Vercel AI frontends, and MCPToolset gains read_resource_contents. MCPToolset.call_tool_for_app relays tool calls made from MCP Apps.
The compatibility notes cover BackgroundTools completions, which now carry namespaced origin and tool_call_id metadata. Realtime history, usage and reply waits for GPT-Live and Gemini Live are now taken from the session core, and Skills now honor disable-model-invocation and user-invocable. Fixes in this release include handling of tool_choice in OpenAIResponsesModel, Gemini thought signatures in OpenAIChatModel, and SpendLimits charging partial responses from interrupted streams.
Why it matters if you build agents
Teams that let models act on tool calls can use the new ApprovalReviewer to auto-approve low-risk calls, but should decide which calls qualify before enabling it. Anyone running BackgroundTools, realtime sessions or Skills should review the compatibility notes, since metadata, history and invocation behavior have changed. Spend tracking in SpendLimits now counts the partial response of an interrupted stream.
Get stories like this every weekday
One short email with the day's agent tooling news.