A workflow for evolving both target agents and the RSI system itself.

English · 简体中文

RSI-Flow is a recursive self-improvement (RSI) workflow that can evolve both target agents, such as Pi and DeepSeek Harness, and the RSI system itself. It uses task experience to build and evaluate changes to prompts, workflows, and other editable artifacts. It compares the original and candidate under a fixed evaluation protocol, applies accepted changes, and supports rollback.

Task execution, proposal generation, and evaluation use separate components. RSI-Flow coordinates the improvement process and records its results.

How it works

flowchart TD
    A[1. Define object and protocol] --> B[2. Import experience]
    B --> C[3. Diagnose observations]
    C --> D[4. Build a candidate]
    D --> E[5. Compare versions]
    E --> F[6. Select and verify deployment]
    F --> B

Loading

Each round defines what may change, how candidates will be evaluated, and the available budget.

Integrations

Framework or tool	Supported role
Pi	Native task execution.
DeepSeek Harness	Native task execution.
Harbor / Terminus-2	Terminal-Bench task execution and result import.
Codex	Proposal generation.

Quick start

Use Python 3.10 or newer on macOS or Linux. From the repository root:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

python examples/run_demo.py --workspace .rsi/quickstart

The local example covers task collection, evaluation, deployment, and rollback. It needs no API key and writes results to the specified workspace. Use a new workspace path for each run.

Code structure

Code	Responsibility
Engine	Coordinate the six-stage improvement process.
Extension contracts	Define interfaces for task execution, proposals, and evaluation.
Integrations	Connect agent frameworks and task environments.
Examples	Runnable examples and configuration files.
Tests	Core behavior and regression checks.

Run the tests

python -m pip install -e '.[test]'
python -m pytest tests

License

RSI-Flow is licensed under the MIT License.