Pydantic AI is a Python framework for building agents, created by the team behind the Pydantic library. The goal is to bring the same type and validation discipline used in Python APIs to agent code.
What it is for
- Building agents with structured, schema-validated outputs
- Exposing Python functions as agent tools with validated typing
- Swapping models without rewriting the agent, including OpenAI-compatible providers
- Testing and evaluating agents with deterministic cases and test models
How it works
- The agent is declared with a model, instructions and typed tools
- Responses can be forced into a Pydantic model, removing fragile parsing
- Dependency injection hands context, connections and configuration to tools
- It supports MCP, agent graphs and stateful flow execution
- Observability uses OpenTelemetry and integrates with Pydantic Logfire
Availability
- Open-source library under the MIT licence, installed with
pip install pydantic-ai - Official documentation with guides, examples and API reference
Points to note
- It is a code framework: it requires Python and programming knowledge
- Version 2.x ships releases quickly; review the changelog when upgrading
- Model choice changes cost and quality; test with real data