Simple yet effective methods for AI4AI.
Simple Agent Lab is an open AI research collective. We are working toward a future where AI can help build AI, while the methods behind that progress remain understandable to humans.
AI will increasingly write code, generate data, run experiments, evaluate results, and improve models and systems. Improved AI can then help build the next generation, forming a recursive loop of improvement. We believe this future should remain simple, transparent, and verifiable so that people can understand how progress happens and participate in deciding what comes next.
We explore:
- AI4AI — using AI to build, evaluate, and improve AI systems.
- Self-Improving Systems — systems that learn from evidence and become more capable over time.
- Recursive Self-Improvement (RSI) — the path from individual improvement loops toward systems that can improve how they improve.
Our direction is AI4AI. Our path is Recursive Improvement. Our principle is Simple.
Simple does not mean unsophisticated. Simplicity is what allows humans to keep understanding and participating in a future built with AI. We want people to see how a method works, verify why an improvement happened, and help decide where it goes next.
Research is better when shared. We aim to share code, experiments, evaluations, and failures so others can inspect, reproduce, and build on what we learn.
- AutoTrainess — teaching language models to improve language models autonomously through a training-specialized Agent–Computer Interface. [Paper]
- Simple Long Horizon Agent — a minimal, understandable agent loop for real long-horizon work. Our members have studied or worked at ByteDance, UC Berkeley, Tsinghua University, Shanghai Jiao Tong University, and Tongji University. All members participate in a personal capacity. See our institutional affiliation disclaimer.
This repository contains the static website for Simple Agent Lab. It uses plain HTML, CSS, and JavaScript and is deployed directly with GitHub Pages.
To preview it locally:
uv run python -m http.server 3000Then open http://localhost:3000/.