About The Improvement Loop
12 illustrated essays on self-improvement as a deployment property — written from the July 2026 literature, every claim traced to its paper.
What this series is
In mid-2026, self-improvement stopped being a training-time trick and became something production systems do: harnesses that revise themselves, skill libraries that compound, memory that trains, and evaluators that must evolve alongside the agents they judge. This series maps that shift substrate by substrate — the working loops, their named failure modes, and the verification that bounds all of it. Every essay closes with a Loop Card: signal, update surface, guardrail, and one check a team can run this quarter.
Where it sits in the family
Harness Engineering (Series 02) is the field guide for building agents today; The Long Bet (Series 03) places the three-to-ten-year bets; and the earlier Continual Intelligence (Series 01) is the cited foundation. The Improvement Loop is the deployment companion: what happens after the agent ships, when the loop starts running on live experience.
How it was made
The corpus is a curated July 2026 collection of 93 papers, indexed, classified, and scored through a staged pipeline with human approval gates at the tag-set and series-design decisions. Drafting followed a strict fact-integrity protocol: abstracts as primary sources, first authors verified on page one of each PDF, every number traced to the paper that contains it, and a battery of verification checks run on every essay before it was marked done. Where the collection lacked a paper's canon, the essays say so rather than cite from memory.
Where to start
Front to back is the intended order — The Self-Improving Stack defines the frame and the notation everything else inherits. If you run agents in production today, the Loop Cards in essays 3, 5, 7, and 8 are the fastest value.