RSIHub
A research framework for principled agent self-improvement under frozen evaluators and declared mutation boundaries, recording verifiable lineage to make it rep
What is RSIHub?
A research framework for principled agent self-improvement under frozen evaluators and declared mutation boundaries, recording verifiable lineage to make it reproducible and auditable.
Overview of RSIHub
What it is
RSIHub is a file-based evolution framework for evaluator-driven learning, reproducible candidate lineage, and controllable modification. It gives an agent a controlled way to improve itself by running candidate agents against a fixed evaluator, keeping the provenance of every generation — what changed and why — and carrying verified improvements forward while the evaluator itself stays untouched. The framework is designed for agent builders and researchers to improve prompts, skills, harnesses, and agent code in a reusable experiment workspace, and to compare evolution strategies under fixed evaluation and mutation boundaries.
Key features
- Provenance tracking: every candidate agent is connected to scores, artifacts, archive records, and Git lineage.
- Composable loop: every recipe follows the same seven-phase loop: select → rollout → analyze → mutate → guardrail → evaluate → absorb.
- Framework-owned evaluation: evaluation is never a selectable operator; it stays framework-owned to ensure inspectability.
- Nine named stages: five required (select, rollout, mutate, gate, record) and four optional (analyze, validate, novelty, reflect).
- Mutable surface declaration: each recipe declares its mutable paths; the evaluator, archive stamps, and framework code stay outside that surface.
- Experiment viewer:
evolve viewserves as a read-only view of one generated workspace, showing experiment health, generation progress, modifications, canonical performance, and paginated trial outcomes. - Agent-driven control: an outer agent can choose research action sequences through
evolve agent, with continuous sessions, versioned methods, and opt-in Docker isolation.
Use cases
- For agent builders: improve prompts, skills, harnesses, and agent code in a reusable experiment workspace.
- For researchers: compare evolution strategies under fixed evaluation and mutation boundaries.
- Skill evolution: improve a Skill as a complete package — instructions, references, and validation scripts evolve together while a frozen evaluator keeps the comparison honest.
- Controlled co-evolution: co-evolve the target and selected evolution policy using recipes like
hyperagents.
Getting started
Run one of the supported recipes against the shared, content-pinned Terminal-Bench 2.0 subset with the following scripts:bash./scripts/setup_terminal_bench.sh./scripts/run_recipe_demo.shFor browsing an experiment, use:bashevolve view /path/to/experimentSee the quick start guide for prerequisites, credential setup, supported recipe values, and launcher overrides.
License and pricing
RSIHub is licensed under Apache-2.0 and is open source. No pricing information is provided beyond the open source license.
Based on the project's public README and repository metadata.