Skip to content
Back to selected work

Scar Tissue

An agent harness that turns each failure into a tested change to its own rules.

Design Engineer & Product Designer · 2026 · Hackathon build

The Scar Tissue demo after order #1: two orders were made, one incident recorded, one of three candidate fixes passed, and policy v2 is live.

Project overview

My role
I scoped the idea, designed the demo and every screen, and directed two coding agents: Claude Code on the interface, Codex on the engine.
My team
No other people. Two coding agents, with one typed state file as the contract between them.
What shipped
A working harness on MongoDB Atlas, a demo page anyone can try to break, a held-out benchmark, and a one-minute video.
What changed
On 11 held-out tasks the same model went from 36% to 94% correct, and duplicate orders fell from 21 to none. Only the harness changed.
What I can share
All of it. The code is public, the demo replays a recorded run, and every number comes from a dated LangSmith experiment.

Read the full case study

Email me