Build a multi-step AI assistant where each step is a separate, testable agent you can swap out independently.
Add web search or file handling to an AI app by downloading pre-built tools from the Atomic Forge catalog.
Create AI pipelines you can audit and debug in production because all logic is plain Python, not autonomous black-box decisions.
Connect to OpenAI, Anthropic, Groq, or Gemini using the same code structure without rewriting your app.
Install via pip with optional extras for non-OpenAI providers (e.g. Anthropic, Groq, Gemini). Requires an API key for your chosen AI provider. Pre-built tools available via the Atomic Assembler CLI.
Atomic Agents is a Python framework for building AI-powered applications by connecting small, focused components together. The core idea is that each piece of an application, whether it is an agent, a tool, or a context provider, does one thing and does it well. You combine these pieces to create more complex behavior, similar to assembling building blocks, rather than writing one large tangled script. Each agent in the framework is defined by three things: a system prompt that tells it how to behave, an input schema that describes what it receives, and an output schema that describes what it returns. The schemas are defined using Pydantic, a Python library for data validation, which means you get type checking and structure guarantees on every message that flows through your application. Context providers are a separate concept that lets you inject dynamic information, like the current date or a user profile, into the agent's instructions at runtime without hardcoding it. The framework is built on top of a library called Instructor, which handles the communication with AI providers. Atomic Agents supports OpenAI by default and can be extended to work with Anthropic, Groq, Google Gemini, and others by installing the appropriate extras. You do not need to learn a new AI provider API because the framework abstracts the differences. A command-line tool called Atomic Assembler ships alongside the framework. It lets you download pre-built tools from a catalog called Atomic Forge, which covers things like web search and file handling, so you do not have to build common utilities from scratch. The design philosophy behind Atomic Agents is to keep control in the developer's hands. All the logic and flow is written in ordinary Python code, so you can apply standard software engineering practices like testing, version control, and code review. This is a deliberate contrast to frameworks that rely on autonomous agents making their own decisions, which can be difficult to debug or audit in production environments.
← brainblend-ai on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.