simple-agent-lab/RSIHub

What it solves

AI agents often need to improve themselves, but doing so blindly can be risky. RSIHub gives agents a controlled way to evolve by testing candidate variants against a fixed evaluator. It keeps a full record of every change (what changed and why) and only accepts improvements that are verified, so the process is reproducible and safe.

How it works

RSIHub runs a repeated loop: select a parent candidate from the archive, run it on training tasks to collect evidence, analyze that evidence to decide what to modify, mutate specific parts (such as prompts, skills, or agent code) while respecting declared limits, guard the changes against going beyond the allowed surface, evaluate the modified candidate with the frozen evaluator, and if the score improves, absorb it as the new parent. Every step is recorded in a repository-like lineage, and all execution happens through Harbor, which provides container isolation, trajectory capture, and per‑task verification.

Who it’s for

Agent builders who want to improve prompts, skills, or agent code in a systematic and reusable way, and researchers who need to compare different evolution strategies under fixed evaluation and mutation boundaries.

Highlights

• Framework‑controlled evaluation: the evaluator is never mutated, so comparisons stay honest.
• Composable recipes: strategies like hill climbing, AHE, GEPA, Hyperagents, and A‑Evolve can be plugged in without writing code.
• Full provenance: every generation is linked to scores, artifacts, archive records, and Git commits.
• Built on Harbor for container execution and per‑task verification, with an optional viewer to inspect runs.