Open research book · 2026 edition
Recursive self-improvement, examined as an engineering discipline.
A practical guide to agents that learn from failure, change their own systems, and produce evidence that the new version is genuinely better.
01 / Reading path
Read the chapters
The State of Recursive Self-Improvement
Map the field to a common loop and locate the evaluation problem at its center.
Build Your First Agent Improvement Loop
Turn a repeated failure into a small system you can run, measure, and inspect.
Anatomy of a Recursive Self-Improvement System
Separate the worker, writer, judge, and gate; then improve from failure traces.
02 / Research index
Notes from the active literature
Each note asks the same questions: what changed, who judged it, what improved, and which conclusion the evidence does not yet support.
autoresearch
A minimal autonomous research loop over one editable training file.
ReasoningBank
Distilling useful strategies from successful and failed trajectories.
LIFE-HARNESS
Adapting the interface around a model from execution failures.
SIA
Co-optimizing the agent harness and underlying model weights.
CL-BENCH
Measuring whether an agent actually learned beyond its base model.
HarnessX
Treating the agent harness as a composable, evolvable object.