Capability · Framework — agents

Pydantic AI

Pydantic AI is an agent framework designed to feel like FastAPI for LLMs. It uses Pydantic models for structured inputs and outputs, provides dependency injection for tools and context, supports streaming and multi-turn conversations, and ships with first-class OpenTelemetry tracing and a built-in evals harness. It's model-agnostic and works with OpenAI, Anthropic, Google, Groq, Mistral, and Ollama.

Framework facts

Category: agents
Language: Python
License: MIT
Repository: https://github.com/pydantic/pydantic-ai

Install

pip install pydantic-ai

Quickstart

from pydantic_ai import Agent
from pydantic import BaseModel

class Answer(BaseModel):
    city: str
    population: int

agent = Agent('anthropic:claude-opus-4-7', output_type=Answer)
result = agent.run_sync('Capital of India and its population?')
print(result.output)

Alternatives

Instructor — structured outputs only
LangGraph — lower-level orchestration
OpenAI Agents SDK — OpenAI's equivalent
Mirascope — typed LLM toolkit

Frequently asked questions

Why pick Pydantic AI over LangChain?

Pydantic AI is smaller, typed end-to-end, and follows the FastAPI design idioms. If you want strict typed I/O, dependency injection, and a minimal footprint rather than LangChain's large ecosystem, Pydantic AI is a strong fit.

Does it support tool calling and multi-agent?

Yes. Tools are declared as typed Python functions. You can compose multiple agents, stream outputs, run structured validations, and retry on validation errors — all with Pydantic types.

Sources

Pydantic AI — docs — accessed 2026-04-20
Pydantic AI — GitHub — accessed 2026-04-20