# Atomic Agents

A lightweight, modular framework for building agentic AI pipelines out of single-purpose, reusable, and composable pieces, built on Instructor and Pydantic.

Atomic Agents treats an agent, a tool, and a context provider each as a small, single-purpose, reusable component that combines predictably with the others, similar in spirit to building with interlocking blocks, so a pipeline can be assembled from pieces rather than one large, hard-to-modify agent class. Building on Instructor and Pydantic lets a developer apply familiar software engineering practices, such as typed schemas, to agent development instead of adopting an unfamiliar framework-specific style.

Installing the Python package also installs its CLI, Atomic Assembler, for downloading community-contributed tools, with agent and pipeline downloads planned. Provider support, including Groq, Anthropic, Gemini, and OpenAI by default, comes through Instructor's provider extras, and the project also ships instructions for AI-assisted development in editors such as Cursor, Windsurf, Cline, and Aider, plus its own agent skills for coding agents.

Version 2.0 is the current major release, and it is distributed under the MIT license.

Website: https://eigenwise.github.io/atomic-agents/