Table of Contents
🌏 中文版
Pydantic AI is a Python agent framework from the Pydantic team. Its defining feature is the type boundary: dependencies and outputs are generic types, while tool inputs and model results are validated with Pydantic schemas.
This fits FastAPI and typed Python backends. An agent can depend on a database client and must return an object the application can use—not text that merely resembles JSON.
Declare dependencies and output together
from dataclasses import dataclass
from pydantic import BaseModel
from pydantic_ai import Agent, RunContext
@dataclass
class Deps:
customer_id: str
class Answer(BaseModel):
summary: str
needs_human: bool
agent = Agent('openai:gpt-5-mini', deps_type=Deps, output_type=Answer)
@agent.tool
async def customer_context(ctx: RunContext[Deps]) -> str:
return f'customer={ctx.deps.customer_id}'The Agent guide groups instructions, tools, output type, dependency type, model, and settings. RunContext passes application dependencies to tools without putting a database or user identity into the prompt.
Validation is not factual correctness
The output guide explains how output types become JSON schemas and validated results. A schema error can trigger a retry; semantic correctness still requires validators, domain rules, or approval. A typed needs_human: false does not authorize a refund.
Tools may use context or be plain functions. Toolsets make integrations reusable and replaceable in tests. For long-running work, the durable execution overview integrates Temporal, DBOS, Prefect, and Restate rather than treating message history as a checkpoint system.
Overall
Pydantic AI fits teams whose outputs enter real Python domain models and whose tools need explicit dependencies. Start with one typed output, one dependency-aware tool, and a fake-model test; deliberately return the wrong schema and observe the failure path. For alternatives, see the agent framework guide.
References
Loading...