News Hub

Pydantic AI agents now checkpoint every model request and tool call to AWS Lambda for resumable execution after failures.

You.com integrates into Pydantic AI as web search and research capabilities with structured citations.

Pydantic AI agents now support real-time speech-to-speech calls across multiple providers.

Guide comparing OpenTelemetry backends in 2026, including metering for applications with LLMs.

Crusoe Managed Inference is now available as a native model provider in Pydantic AI with streaming, tool calling, and structured output support.

StackOne integrates as a capability in Pydantic AI, enabling agent access to SaaS endpoints without custom tool code.

Hack Monty Round 2 sandbox was not breached. Round 3 offers $20,000 bounty on new WebSocket service.

Pydantic AI now provides native Snowflake integration for running governed agents within Snowflake.

Pydantic AI and Logfire enable Airbnb's three-layer evaluation workflow for prompt improvement.

Run Braintrust Python and TypeScript evals against Logfire by changing two environment variables without rewriting the eval suite.

Logfire charges no per-score fee for running evals and provides tools to compare runs and inspect results.

Logfire does not charge per score, unlike Braintrust, enabling higher evaluation coverage at scale.

An agent loop that iterates until goals are met, remembers runs, and self-grades using a calibrated judge.

Pydantic released official skills for Validation, AI, and Logfire for Claude Code, Codex, and Cursor.

DynamicWorkflow feature enables orchestrating multiple Claude agents; used to rewrite Bun from Zig to Rust in eleven days.

MCP Python SDK v2 beta implements the 2026-07-28 spec with multi-round-trip requests and Pydantic Logfire tracing.

Multiple agents can use disposable cloud instances to build and test designs in parallel without using production infrastructure.

A CVE-bump agent runs in fifteen lines on a laptop; ModalSandbox enables production scaling with gVisor containers at five hundred parallel tasks.

Agent-written code arrives faster than humans can review it; LLM self-critics and AST-aware review tools enable faster merge workflows.

Pydantic AI offers three research agent approaches: native WebSearch, ExaAgent for multi-step research, and composed ExaSearch for custom loops.