Self-Improving Agents

What

Agents that improve themselves — design their own prompts/modules (automated agent design), run open-ended research loops, and write better code over time.

Papers

• The AI Scientist (Lu et al. 2024) — fully automated research loop: idea → experiment → paper → (self-)review, at a fraction of human cost. Includes automated reviewing of its own output.
• AlphaEvolve (2025, Gemini) — a coding agent that designs advanced algorithms: evolutionary loop over algorithm ideas, verified by execution on benchmarks; found algorithms that beat human-designed ones.
• CodeMonkeys — scaling test-time compute for software engineering: massively parallel candidate patches + verification selects fixes; improves SWE-bench.
• KernelBench — benchmark for LLMs writing efficient GPU kernels — the frontier of code-gen ability.

Key ideas

• The loop: generate → execute → verify → keep what works → repeat. Execution is the ground truth.
• Self-review is part of the loop (AI Scientist reviews its own papers).
• Evolutionary/selection beats single-shot for hard design problems.

SolOS mapping

• Auto-skill-extraction (adoption #6): mine successful sessions → skill candidates → verify by re-use → keep. The skill system is our "genome."
• Selection loop we can run today: when a brief/script pattern succeeds, promote it; when it fails, demote/patch — our skill_manage discipline formalized as a weekly "evolution pass."
• AI Scientist lesson for the fleet: every deliverable gets a self-review step before sign-off (verifier pass again — it's the common thread).