Mapping RSI: a taxonomy of self-improving agents from weights to harness code

青稞AI · wechat · 2026-09-14

A systematic survey of Recursive Self-Improvement (RSI), organizing existing work into a coordinate system for analyzing new approaches.

Definition: agents update their own state (model and harness) from task trajectories and feedback to improve future performance.

Five evolution targets, with representative work:

• Parameter evolution (SEAL): models generate self-edits for self-training, with RL filtering effective edits; expensive and hard to revert
• Context evolution (PrimeAgent): persistent REPL workbench plus Continual Harness, sub-agent parallelism, no weight updates
• Memory evolution (ReasoningBank): distilling success/failure trajectories into reusable memories, +3.7–6.2pp success on WebArena/SWE-Bench; MaTTS scales test-time attempts
• Skill evolution (TRACE): a SkillBank revised from traces; GPT-5.5 3x-consistent success rate up from 59.9% to 94.5%
• Harness code evolution (SkillSmith): co-evolving skills and tools using ecological cooperation/competition relations

Evolution structures: chains (SkillFlow, Claude Opus 4.6 62.65%→71.08%), trees (Darwin Gödel Machine, SWE-bench 20.0%→50.0%), and graphs.

Original post →

More from coding & agent

• Developer tells agent to 'figure out' the API key itself — and it does — shakoistsLog · 2026-09-21
• Pretty Mermaid: Open-Source Skill Renders Mermaid to SVG/PNG/ASCII Locally, No Browser — tom_doerr · 2026-09-21
• Handing my tax credentials to an AI agent: great UX, unsettling security — zeeg · 2026-09-21
• Coding agent AdaL opens free access to 10 users, inviting harshest criticism — Zachly · 2026-09-21
• Dev spins up a custom Grok Bot "anew" to serve free AI webpages on X — round · 2026-09-21
• JevArena: open-source arena for blind-testing AI judges on LM-as-a-Judge tasks — richie9830 · 2026-09-21

coding & agent channel →