// the find
pydantic/pydantic-ai
AI Agent Framework, the Pydantic way
Pydantic AI is an agent framework from the team behind Pydantic itself — the validation library that already underpins most of the Python LLM ecosystem. It brings typed, dependency-injected agents with structured outputs to Python developers who want something more principled than LangChain without rolling their own abstractions. The target audience is Python developers building production agents who already think in Pydantic models.
The dependency injection model for tools is genuinely well-designed — passing typed deps through RunContext means tools are testable in isolation without mocking global state. Structured output validation with automatic LLM retry on schema failure is the right default behavior that most frameworks leave to the user. The evals system is first-class, not an afterthought, and integrates directly with Logfire for tracking regressions over time. Model provider coverage is unusually wide — the same agent code runs against OpenAI, Anthropic, Bedrock, Ollama, and a dozen more without changes.
The tight integration with Pydantic Logfire for observability is real value but also a soft vendor push toward a paid product — the OTel alternative path is documented but visibly second-class in the examples. The graph support for complex workflows is powerful but adds a distinct programming model that doesn't compose naturally with the simpler agent API, so complex projects end up learning two systems. Durable execution is documented as integrations with Temporal, DBOS, and others rather than a built-in primitive, which means fault tolerance in long-running agents still requires a separate orchestration dependency. The framework is moving fast enough that the changelog shows breaking changes in minor versions, which is a real cost for teams trying to pin a stable production dependency.