It uses just LLM sampling to:
- Autonomously write, test, and push code
- Read GitHub issues and add new features, making PRs in response to them end-to-end
- Resolve git and environment issues and backtrack autonomously
- Perform web research tasks (e.g., gathering niche trivia from Reddit)
Some architectural details. It uses:
- ReAct-based iterative reasoning with tool usage (web search, file operations, code execution)
- Short- and long-term memory management for context
- Various engineering tricks to handle surprising edge cases that arise when going from LLM -> agent
The result is a simple self-contained repo showing people the core logic needed to go from "language model" to "autonomous agent" and unpack the hype/jargon around agents.
It's intentionally minimal and educational, but seeing genuine autonomy arise from simple token sampling was a striking "feel the AGI" moment for me! I welcome contributions from others.