# Pydantic AI

> Typed Python agent loop with a snap-on Coder harness

**Category:** AI Agents  
**Pricing:** open-source  
**URL:** https://ai.pydantic.dev  
**GitHub:** https://github.com/pydantic/pydantic-ai  
**Added:** 2026-09-03

---

## The Hook

Pydantic AI is how the Pydantic team does agents—typed end to end, not another wrapper around the validator.

## What's Good

This is a different product from Pydantic-the-validator already on the Python stack. Core is a typed agent loop: structured output, typed deps, typed tools, model as a string swap (OpenAI, Anthropic, Gemini, Bedrock, Groq, Ollama, and more). Pydantic AI Harness snaps on Coder—workspace-rooted files, allowlisted shell, repo orientation, planning, sub-agents, context management—as a regular combined capability you can use whole or take apart (FileSystem, Shell, RepoContext, Planning). Same agent runs as CLI (clai), web chat, realtime voice, or a durable Temporal/DBOS/Prefect job. MIT.

## The Catch

ACP (Agent Client Protocol, for editors like Zed) is experimental: it lives under pydantic_ai_harness.experimental.acp, may change or be removed in any release without a deprecation period, and importing it emits HarnessExperimentalWarning. Coder is a capability you assemble, not a turnkey SWE agent like OpenHands. Logfire is the convenient observability path; OpenTelemetry works with any backend. Do not confuse this gem with Pydantic validation.

## Verdict

The typed Python agent SDK. Not the validator. ACP is still experimental.