A field guide to agents that rewrite themselves

What it does

Awesome RSI is a curated index of research on Recursive Self-Improvement — agents that do tasks, learn from trajectories and feedback, update their own state, and use that updated state next time. That state might be model parameters, harness code, context, memory, or skills, and the whole collection is organized around exactly that question: what is the agent improving? It ships with a companion website offering filters, citation graphs, and an illustrated intro guide (in English and Chinese).

The interesting bit

The taxonomy is the real contribution. Splitting RSI work by artifact — parametric (model weights) vs. non-parametric (harness code, context, memory, skills) — turns a fuzzy buzzword into a navigable design space. The accompanying survey adds a nine-dimension taxonomy plus lessons borrowed from traditional deep learning.

Key highlights

• 113 method papers, 33 benchmark papers, and 3 systems, with counts auto-generated into the README
• Paper map tree grouping methods by RSI artifact, with example papers on each branch
• Interactive citation graphs on the website, switchable between methods and benchmarks
• Companion survey: “From Existing Methods to New Research Ideas: Two Perspectives on Recursive Self-Improvement”
• Works can appear under multiple artifacts; inclusion covers both explicit self-modification and experience-driven improvement during long-running tasks

Caveats

• The overwhelming majority of entries are arXiv preprints — a fast-moving field indexed at preprint velocity, so treat individual results with the usual caution
• Only 3 systems made the cut, so if you want runnable code rather than reading material, the pickings are thin

Verdict

Essential bookmark material if you’re researching or building self-improving agents — it’s the map of the territory. If you just want a working agent framework today, the three systems listed are where to start and stop.