Agent delivery harness
How AI agents ship our software, with every change checked before it lands
Git / TypeScript / Node.js
The brief
The problem
AI agents write code quickly and report success confidently, even when the work is incomplete. Run several at once and they collide in the same files. Let one grade its own work and it passes.
Our approach
We built a harness that checks the agents' work. Work starts as an item on a durable board with written acceptance criteria. Each agent gets its own isolated copy of the repository, so parallel work cannot collide. When it finishes, a deterministic gate (build, tests and checks that do not care what the agent claims) decides whether the work is real. A reviewer model from a different AI vendor then reads the change. Only work that clears both gates lands, and a person owns what ships.
Features
What it does
- 01
Durable work board with acceptance criteria
- 02
One isolated worktree per agent
- 03
Deterministic build-and-test gate
- 04
Review by a second AI vendor
- 05
Human-owned landing
Results
What happened
Runs every day on our own products, including this website
Parallel agents work in isolated copies and never share a working tree
Nothing lands on an agent's own say-so: a deterministic gate decides
A second model from another vendor reviews each change
The board survives restarts and machine changes, so no work is lost mid-flight
Working on something similar?
Tell us what you are building or fixing.