RSIHub¶
Build agents that improve — and keep the evidence.
RSIHub is a file-based framework for running agent-evolution experiments without rebuilding candidate snapshots, evaluation, lineage, and reporting for every method. It provides composable recipes inspired by A-Evolve, AHE, GEPA, Hill Climb, and HyperAgents.
Start here¶
Run an experiment¶
| Step | Guide |
|---|---|
| 1. Install RSIHub and verify the local setup | Getting started |
| 2. Configure credentials and the host runtime | Environment Variables |
| 3. Initialize a workspace and launch an experiment | From recipe to experiment |
| 4. Check, monitor, and recover a real run | Running RSIHub reliably |
Build your own method¶
| Goal | Guide |
|---|---|
| Compose a custom experiment configuration | Creating a custom recipe |
| Choose stages and built-in variants | Operator overview |
| Configure the editing agent and isolation model | Meta-agent execution |
| Run trusted local tasks without Docker isolation | Local Harbor environment |
For the framework model and vocabulary, see Framework design and Terminology.
What RSIHub keeps fixed¶
Each experiment is a separate Git repository. Generation tags identify exact
candidates, archive.jsonl records stamped outcomes, and the evaluator stays
outside the candidate's mutable surface.
The mechanism enforces three core rules:
- Scores and statuses are written by the mechanism, not workspace operators.
- Canonical evaluation runs on clean candidate snapshots against a frozen evaluator.
- Reports are recomputed from stamped archive records rather than mutable operator claims.
See the design guide for the complete model and invariants.