Repolode
Submit
simple-agent-lab

simple-agent-lab/RSIHub

A Python framework for evaluator-driven agent self-improvement, recording lineage for every candidate agent generation.

View on GitHub
Stars
341
Forks
19
Open issues
7
Last commit
11 days ago

What is it?

RSIHub runs candidate AI agents against a fixed evaluator, keeps a record of what changed in each generation and why, and carries verified improvements forward while the evaluator stays untouched. Each evolution strategy is expressed as a recipe that selects parent agents, analyzes execution traces, proposes edits within a declared mutable surface, and decides which candidates are admitted.

Every trial runs on Harbor, which provides container execution, trajectory capture and per-task verification. Candidates are scored from clean snapshots, and results are recomputed from stamped archive records and Git generation tags. The mutable surface can cover prompts, skills, harnesses and target agent code, and recipes such as hill_climb, aevolve, ahe, gepa and hyperagents are included.

The project also ships an experiment viewer, `evolve view`, that inspects generated workspaces read-only, showing generation progress, modifications, scores and trial outcomes. It is an active research prototype.

Key features

  • 01Evaluator stays frozen, so candidate agents cannot change the scoring contract.
  • 02Each recipe declares the target and operator paths that may be modified.
  • 03Candidates are scored from clean snapshots of the agent code.
  • 04Lineage is recorded in stamped archive.jsonl entries and Git generation tags.
  • 05Trials execute on Harbor with container execution and trajectory capture.
  • 06Read-only experiment viewer for inspecting generations and trial outcomes.

Use cases

  • Improving prompts, skills or harness code of an agent under a fixed evaluation suite.
  • Comparing agent evolution strategies under identical evaluation and mutation boundaries.
  • Keeping a reproducible, auditable record of how an agent's versions were derived.

Quick start

  1. 1

    Set up the shared Terminal-Bench 2.0 subset

    ./scripts/setup_terminal_bench.sh
  2. 2

    Run a supported recipe demo

    ./scripts/run_recipe_demo.sh
  3. 3

    Review prerequisites and credential setup in the quick start guide

  4. 4

    Browse a generated experiment workspace in the read-only viewer

    evolve view /path/to/experiment

Things to know

  • Requires Python 3.12 or newer.
  • Benchmark runs use GPT-5.4 and a Codex mutate operator, so you will need access to those paid model APIs.

In collections

Open-source AI agent frameworks & apps

Topics

#agent#self-improvement#llm-agents#evolutionary-search#reproducibility#evaluation#python#agent-evolution#agentic-ai#ai

FAQ

What is RSIHub?
A Python framework for evaluator-driven agent self-improvement, recording lineage for every candidate agent generation. RSIHub runs candidate AI agents against a fixed evaluator, keeps a record of what changed in each generation and why, and carries verified improvements forward while the evaluator stays untouched. Each evolution strategy is expressed as a recipe that selects parent agents, analyzes execution traces, proposes edits within a declared mutable surface, and decides which candidates are admitted.
Is RSIHub free to use? What license does it have?
RSIHub is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is RSIHub still maintained?
Actively maintained. The last commit to RSIHub was on Oct 27, 2026.
What can I use RSIHub for?
Improving prompts, skills or harness code of an agent under a fixed evaluation suite. Comparing agent evolution strategies under identical evaluation and mutation boundaries. Keeping a reproducible, auditable record of how an agent's versions were derived.
How do I install RSIHub?
Set up the shared Terminal-Bench 2.0 subset: ./scripts/setup_terminal_bench.sh Run a supported recipe demo: ./scripts/run_recipe_demo.sh Review prerequisites and credential setup in the quick start guide Browse a generated experiment workspace in the read-only viewer: evolve view /path/to/experiment
Where can I get RSIHub?
The source code and installation instructions are on GitHub: https://github.com/simple-agent-lab/RSIHub

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects