GenericAgent: The OS-Level Assistant That Learns to Stop Thinking

How a 90-line core loop uses skill crystallization to turn messy autonomous exploration into a permanent, high-precision library of system tools.

• View on GitHub • More from louislva

A mechanical arm presses a chaotic scribble into a geometric crystal mold, representing the crystallization of messy AI reasoning into permanent tools.

Since 2022, I’ve had one very clear vision: build the most complete resource for engineers who want to work with real AI systems. Not prompt tricks. Not toy notebooks. Not “look, it calls a tool.” Real systems. The kind that survives production.

Louis-François Bouchard, Co-creator, Towards AI · Stop Building Agent Demos

Key Takeaways

GenericAgent refines messy autonomous reasoning into permanent bash scripts through a process called crystallization.
A minimalist 90-line core loop reduces the surface area for hallucinations and framework bloat.
The agent builds its own capability library by saving successful execution paths as reusable local tools.
This system prioritizes organic growth over the complex dependency chains found in traditional orchestration frameworks.

The End of Ephemeral Reasoning

The agent era is defined by two extremes. On one side are massive orchestration frameworks. On the other are ephemeral chat loops that forget how they solved a problem the moment the process ends. GenericAgent introduces a third path called crystallization.

It is a minimalist 90-line loop that does not just execute tasks. It saves successful execution paths as permanent, one-line skills. The agent builds its own library, moving from a stumbling autonomous solver to a precise tool-user over time.

Inside the 90-Line Engine

The core engine strips away the bloat of traditional frameworks. The entire execution loop lives in a single Python file. This minimalism is not just an aesthetic choice. It is a reliability strategy. Less code means a smaller hallucination surface and a tighter feedback loop.

The 90-line core loop connecting the LLM directly to system interfaces.

Crystallization: From Trace to Tool

When GenericAgent succeeds at a complex task, it does not simply exit. It refines the messy history of its actions into a clean script. A chaotic search for a file becomes a concise bash script stored in a local skills folder.

A split illustration showing a hiker lost in a dense forest on the left, and a paved road through the same forest on the right, illustrating the shift from messy exploration to a permanent tool. — Crystallization turns a wandering execution trace into a direct, reusable path.

The Agent That Built Itself

The project serves as a foundational reference for developers moving from toy notebooks to reliable production systems. The initial repository was actually bootstrapped by the agent itself, handling Git commits and environment setup autonomously.

Portrait of Louis-François Bouchard rendered in black ink on a white background.

The Thin Framework Advantage

GenericAgent fills a specific gap for users who want zero deployment overhead and organic growth. It rejects the complex dependency chains of industrial agents in favor of a lean seed that grows capabilities through use.

Framework	Architecture	Memory Strategy	Best For
GenericAgent	90-line procedural loop	Crystallized scripts	Local OS autogrowth
XAgent	Tripartite (Dispatcher, Planner, Actor)	Hierarchical state	Enterprise task breakdown
OpenInterpreter	Natural language terminal	Ephemeral context	Interactive coding sessions