Type-Safe Agents: Use Pydantic AI to Get Validated Python Objects Instead of Raw LLM Strings

Pydantic AI wraps your LLM call, sends your Pydantic model's JSON schema, validates the response, and retries on failure — you get a typed Python object back, never a raw string to parse.

What you'll be able to do after this:

• Declare a BaseModel as your agent's output_type and receive a fully-validated Python object — with IDE autocomplete and type checking — after every call
• Write agents that automatically retry when the LLM returns malformed JSON, without any try/except boilerplate
• Swap between Anthropic Claude, OpenAI GPT, and a local Ollama model by changing one string in your agent constructor

Walk-through

Install:

pip install "pydantic-ai[anthropic]"

Define your output model and run:

from pydantic import BaseModel
from pydantic_ai import Agent

class BugReport(BaseModel):
    severity: str          # "low" | "medium" | "high" | "critical"
    component: str
    root_cause: str
    suggested_fix: str

agent = Agent(
    "anthropic:claude-sonnet-5",   # swap: "openai:gpt-5.6" or "ollama:llama3"
    output_type=BugReport,
)

result = agent.run_sync(
    "The login button does nothing on mobile Safari 17. "
    "Stack trace: TypeError: Cannot read properties of undefined (reading 'submit')"
)

# result.data is a validated BugReport — not a string
print(result.data.severity)       # e.g. "high"
print(result.data.suggested_fix)  # type-checked, IDE-autocomplete-ready

Pydantic AI translates your model's fields to a JSON schema, sends it alongside your prompt, validates the LLM's JSON response, and retries automatically (configurable max_retries) if validation fails. Add tools by decorating plain Python functions with @agent.tool — they get included in the schema automatically. For async workloads, replace run_sync with await agent.run(…).

Provider switching is one string change: "anthropic:claude-sonnet-5" → "openai:gpt-5.6" → "ollama:llama3". The framework supports 20+ providers. Full output patterns (streaming, union types, lists of models) are in the Output docs.

Sources: Output — Pydantic AI Docs · Pydantic AI overview · YouTube: Build 100% Reliable AI Agents with Structured Output · Building AI Agents in Python with Pydantic AI — Machine Learning Mastery