Durable Agent Harness

Public lab note from Dynamic Devices on building a durable operating layer around AI coding agents.

Live site: https://dynamicdevices.github.io/durable-agent-harness/

Why this exists

Frontier models are powerful and unreliable in the same breath. A harness — skills, rules, curated memory, signed patterns, and human gates — is how engineering work stays fast without becoming careless.

This repository is a privacy-safe map of one team’s journey since serious hands-on AI work began (2025-06-06): tenure clocks, the stack, a capability view, patterns worth stealing, and honesty about measurement.

Explore

Path	What
docs/	GitHub Pages site
content/	Source JSON for the site (synced into docs/content/)
PRIVACY.md	Publication bar — what must never ship
tests/	Playwright UX + rendered-text privacy checks

Develop

npm install
npx playwright install chromium
npm test          # privacy gate + Playwright (desktop + mobile)
npm run serve     # http://127.0.0.1:4173

Edit files under content/, then npm run sync (also run automatically by npm run privacy).

What you will not find here

Private chats, family details, credentials, lab network data, client identities, or competitive product plans. See PRIVACY.md.

License

Source and site content: MIT.
Dynamic Devices mark: retained by Dynamic Devices Ltd; used here for attribution on an official public lab note.