Atomic Agents: building AI agents like LEGO bricks

The idea: single-purpose building blocks

The creator of Atomic Agents compares it to LEGO: each component does one thing, can be reused, combines with others and produces consistent results. In practice, every agent declares an input schema and an output schema with Pydantic, and the Instructor library forces the large language model to answer in that format (structured output). To chain two agents, you match one agent’s output to the next agent’s input.

Around these agents you get a system prompt generator (background, steps, output instructions), a chat history and “context providers” that inject dynamic information such as search results.

What it’s good for

Atomic Agents fits controlled processing pipelines: an assistant that always suggests follow-up questions, a pipeline that searches the web and then summarizes, an agent that extracts specific fields from a document. The “atomic” CLI downloads ready-made tools (web search, calculator, PDF reader, YouTube transcripts and more) as source code you can edit. An MCP connector lets you use tools from an external server, and every run can be traced with OpenTelemetry.

Getting started

Install atomic-agents with pip, add the Instructor extra for your provider if needed, then define an output schema, an agent configuration and a model client. The repository includes quickstart examples and complete projects, and the documentation also publishes files designed to be read by a coding assistant.

Strengths and limitations

Strengths: explicit, testable code, typed outputs, little hidden “magic”, many model providers and an MIT license.

Limitations: Python only; the approach requires writing more schemas than more automatic frameworks; the project relies largely on a single maintainer, and version 2.0 broke compatibility with the previous one.

Alternatives

Pydantic AI shares the same typing philosophy, smolagents is just as light but more autonomous, and Langroid is built around agents that exchange messages.