The Lab · Research & Development
OverkillOS: the operating method behind Harmony's agent fleet
The operating method behind Harmony's agent fleet.
The problem
AI coding agents start every session with no memory. They re-learn the project, redo work, break what worked yesterday, and report "done" with nothing to show for it. Rules files help, but they don't give an agent memory, a plan or a definition of done.
What we built
- A method, written as files. Tiered memory: a ~300-token QUICKSTART read at every boot, MEMORY on demand, ARCHIVE for history. There is a global store across projects and one per project. A boot protocol, research-to-PRD templates per area, and execution plans made of gates. A gate closes only with an evidence row, and nothing that worked may regress.
- Host playbooks for Cursor, Antigravity and OpenClaw.
- A local control-plane app (Next.js, 21 screens) that shows every project's state, memory, diagnostics, template sync and gates in one place.
What it does now (2026-09-23)
Template v1.4.0 (last commit 2026-04-01) is MIT licensed. The docs are published at https://overkillos.vercel.app. The kit and the app are in early access; nothing is sold. The app runs locally only.
Outcome
Harmony runs its own projects on the method. We publish no productivity figures because none have been measured.
Built with the Harmony fleet
Written and refined by Harmony's founder with the agent fleet it now runs; the docs site was built during the 2026-09-23 product launch.