pydantic-ai

Pydantic AI is a Python library for building applications and agents that use generative AI models, the kind of AI that produces text, answers questions, or takes actions based on instructions. An agent here means a program that wraps a language model, gives it tools, and runs it in a structured way. The library is built by the team behind Pydantic, a popular Python tool that checks the shape of data, and its stated goal is bringing the FastAPI feel to this newer category of AI apps.

Conceptually, you create an Agent object, tell it which underlying model to talk to, give it instructions, and then call methods like run_sync with a prompt to get a result. From there you can add tools the agent can call, dynamic instructions, structured outputs the model must conform to, and composable capabilities like web search or step-by-step thinking. The README says the library is model-agnostic and supports many providers including OpenAI, Anthropic, Gemini, DeepSeek, Grok, Cohere, Mistral, and others, with a path for custom ones. Other listed features include type-safe APIs for IDE help, an evals system for testing agent quality, observability through Pydantic Logfire or any OpenTelemetry backend, support for Model Context Protocol and Agent-to-Agent interop, streamed structured outputs, human-in-the-loop tool approval, durable execution, and graph definitions for complex flows.

You would use Pydantic AI when writing a Python backend that needs an LLM-powered feature or agent and want validation, type safety, observability, and a single API across many providers. The full README is longer than what was provided.