Pydantic Ai

PydanticAI brings the Pydantic team's type-safety philosophy to agent development — structured outputs are Pydantic models, dependency injection is first-class, and the testing utilities make agent unit testing practical. Model-agnostic (OpenAI, Anthropic, Gemini, Ollama). Particularly compelling for FastAPI-stack Python teams that already trust Pydantic for data validation. Clean, explicit API design. Confidence is docs-derived.

Python SDK: pip install pydantic-ai. @agent.tool decorated functions define tools. agent = Agent(model, result_type=MyPydanticModel, system_prompt=..., deps_type=MyDeps). result = await agent.run(user_prompt, deps=MyDeps(...)). result.data is a validated Pydantic model. Dependency injection via Depends pattern (familiar to FastAPI users). Streaming: await agent.run_stream(...) for streamed responses.

No PydanticAI API key — open-source library. Model provider keys configured per-model (OpenAIModel(api_key=...), AnthropicModel(api_key=...), etc.). Dependency injection handles secrets cleanly via deps pattern. Open-source MIT license.

Agent run failures surface as ModelRetry (retried automatically) or UnexpectedModelBehavior exceptions. result_type Pydantic validation failures: ValidationError surfaced as agent error — configure retries. Testing: TestModel for deterministic tests without real LLM calls. Active GitHub (pydantic/pydantic-ai) with rapid release cadence.

ai.pydantic.dev provides clean, well-organized documentation — quickstart, agent definition, tools, structured results, dependency injection, streaming, and testing guide. Getting started: pip install pydantic-ai, first typed agent in under 10 minutes. Strong type annotations throughout the docs. Community via Pydantic Discord and GitHub.