Eigenwise/atomic-agents

What it solves

Atomic Agents provides a lightweight, modular framework for building agentic AI pipelines. It addresses the lack of control and predictability often found in autonomous multi-agent systems, enabling developers to create AI applications that produce consistent, reliable outputs aligned with specific business objectives using standard software engineering principles.

How it works

The framework is built on Instructor and Pydantic, treating AI components like "LEGO blocks." It decomposes agents into single-purpose, reusable, and composable parts. Each agent consists of a system prompt, an input schema, an output schema, and a history mechanism.

Key mechanisms include:

• Context Providers: Classes that inject dynamic information into the system prompt at runtime.
• Schema Alignment: By aligning the output schema of one agent with the input schema of another (or a tool), components can be chained together or swapped effortlessly.
• Atomic Forge: A collection of pre-built tools (e.g., search, scrapers, calculators) that can be integrated into agents.

Who it’s for

It is designed for developers who want to build predictable and maintainable AI agents and pipelines using Python, applying familiar software development best practices to LLM applications.

Highlights

• Strong Typing: Uses Pydantic for strict input and output validation.
• Modular Architecture: Components are single-purpose and reusable across different pipelines.
• Dynamic Context: Runtime injection of data via Context Providers.
• AI-Assistant Ready: Provides dedicated project instructions and LLM-ready documentation bundles for tools like Cursor and Claude Code.
• Extensible Tooling: Includes a CLI (Atomic Assembler) for managing and downloading a library of specialized tools.