PydanticAI

Architecture Overview

PydanticAI centers around an Agent class parameterized by a dependency type and a result type. The agent uses a system prompt (which can be dynamic via decorators), tools (registered via decorators with dependency injection), and a model. When run, it enters a loop where the model generates responses validated against the result type using Pydantic. Dependencies are injected into tools at runtime via a RunContext, enabling clean separation of concerns and testability.

When to Use PydanticAI

• Type-safe agent outputs with guaranteed schema compliance
• Structured data extraction from unstructured text
• API-driven agents that need validated responses
• Applications requiring dependency injection and testability
• Production agents where output reliability is critical

Strengths & Weaknesses

Quick Start

from dataclasses import dataclass
from pydantic_ai import Agent, RunContext

# Define dependencies (injected at runtime)
@dataclass
class WeatherDeps:
    api_key: str
    units: str = "celsius"

agent = Agent(
    "openai:gpt-4o",
    deps_type=WeatherDeps,
    system_prompt="You are a helpful weather assistant.",
)

@agent.tool
def get_weather(ctx: RunContext[WeatherDeps], city: str) -> str:
    """Get the current weather for a city."""
    # ctx.deps gives access to injected dependencies
    return f"72°F in {city} (using key: {ctx.deps.api_key[:8]}...)"

# Run with injected dependencies
result = agent.run_sync(
    "What's the weather in Tokyo?",
    deps=WeatherDeps(api_key="wk-abc123def456"),
)
print(result.data)

Features at a Glance

Notable Users

Resources