Pydantic AI

Pydantic AI is a Python agent framework built by the team behind Pydantic, the validation library that already powers the OpenAI SDK, the Anthropic SDK, LangChain, and most of the Python AI ecosystem. It brings the same type safety and developer experience to building AI agents, so every input, output, and tool call is validated before it reaches your application logic. Instead of parsing raw text and hoping the model followed instructions, you define a Pydantic model for the output you want and the framework guarantees the response matches it. That guarantee is what lets developers ship agents to production with confidence instead of duct-taping error handling around unpredictable text responses.

How Pydantic AI works

Pydantic AI treats an AI agent as a typed Python object with a defined input, a defined output, and an optional set of tools it can call along the way. Everything is checked by your IDE and type checker before the code ever runs, not just at runtime.

• Agents: An agent wraps a model, a set of instructions, and an expected output type into a single reusable object. You can call the same agent synchronously or asynchronously depending on your application.
• Structured output: You define the shape of the response as a Pydantic model. If the model’s raw output does not match that shape, Pydantic AI automatically asks it to retry, a process called reflection, until the output validates or the retry limit is reached.
• Function tools: Plain Python functions become tools the agent can call mid-conversation. The LLM reads the function’s type hints and docstring to know what the tool does and what arguments it expects, so you rarely write manual tool schemas by hand.
• Dependency injection: A RunContext object carries external dependencies, like a database connection or API client, into your tools and instructions at runtime. This keeps agent definitions clean and makes them easy to unit test with mock dependencies.
• Capabilities: Reusable bundles of tools, hooks, instructions, and model settings that plug into an agent. Built-in capabilities cover web search, extended thinking, and MCP, and you can build or install your own.
• Observability: Pydantic AI instruments agents with OpenTelemetry and integrates directly with Pydantic Logfire, so every tool call, retry, and token cost shows up in a trace you can inspect after the fact.

What you can build with Pydantic AI

• Customer support agent: Build a support agent that pulls account data through dependency-injected database tools, checks a customer’s balance or order status, and returns a structured risk score alongside its response so downstream systems can act on it automatically.
• SQL generation tool: Give the agent a database schema and let it translate natural-language questions into validated SQL queries, with the output type enforcing that the generated query matches an expected structure before it ever touches your database.
• Data analysis agent: Wire the agent to internal APIs and spreadsheets so it can gather figures from multiple sources, apply business logic, and return a structured report, turning a task that takes an analyst hours into one that runs in seconds.
• RAG-powered research assistant: Combine Pydantic AI with a vector database to retrieve relevant documents, then have the agent synthesize a cited answer with a guaranteed output schema, useful for internal knowledge bases and technical documentation search.
• Multi-step workflow automation: Use Pydantic Graph, the framework’s companion library, to model a complex process, like an approval pipeline or a booking flow, as an explicit graph of steps instead of letting control flow degrade into nested conditionals.
• Durable long-running agent: Pair Pydantic AI with a durable execution backend like Temporal or DBOS to build agents that survive API failures, application restarts, and human-in-the-loop approval delays without losing their place in the workflow.