Knowledge-centric Self-Improvement

A population of disposable agents attempts your own tasks, compares notes on what worked, distills the lessons into reusable knowledge, and seeds the next generation with it.

Get started Paper Blog GitHub

•   Getting started

  ---

  Fresh clone to a solved demo task in one command — then your own tasks and larger runs.

  Run the demo

•   Your own tasks

  ---

  Point KSI at any JSON/JSONL file of tasks — no dataset or loader code needed.

  Bring your own tasks

•   FAQ

  ---

  What is this, when to use it, what it costs, and where results go.

  Browse the FAQ

•   Programmatic API

  ---

  Drive a run from Python with ksi.run(...) — plus the auto-generated API reference.

  API guide

•   Architecture

  ---

  The runtime execution chain and which component owns each database.

  Read the overview

•   Extending KSI

  ---

  Add a benchmark, evaluator, runtime, or improvement strategy — one register_* call each, no core edits.

  Extend KSI

•   Reference

  ---

  The glossary, artifacts & cleanup, and runtime startup performance notes.

  Browse the reference

•   Benchmarks

  ---

  Run the reference benchmarks yourself — data prep, task maps, and run presets.

  See benchmarks

Note

The files under docs/ remain the canonical source; this site renders them plus an auto-generated Python API reference.