doorman is a self-improving harness that sits at a website’s door. One script tag maps the page into tools and exposes them to visiting AI agents through WebMCP document.modelContext when available, a /.well-known/doorman.json manifest, and window.doorman. An agent can call a tool directly with zero inference, or hand a goal to the harness, which decides each step with Jev, a System-1 model, in a few hundred milliseconds, with an LLM only as fallback. The harness config (which tools are exposed, what context the model sees, the confirmation guardrails, the rules) is a versioned document in MongoDB Atlas, and the harness rewrites it: an editor LLM proposes JSON-patch edits, code validates them, and the candidate is benchmarked on held-out goals and kept only if it measures better. Every map, visit, step, recipe and benchmark lives in Atlas, and the site owner gets intent analytics from it: what agents came to do, where the site stopped them, and what they asked for that the site does not offer. Measured today, live on Fly.io. Against an ordinary LLM browser agent on the same goal and fresh cart, doorman reached the goal in 1.6 s with 0 LLM calls versus 6.1 s and 4. Over 23 held-out goals it made 4x fewer LLM calls and was 1.6x faster; the baseline still wins on success rate (100% vs 93%), and neither agent declines impossible goals yet. The evolution loop proposed two candidates; both were benchmarked and rolled back, which we show as regression protection rather than a gain. Everything was built today, and every number in the video is measured and stored in Atlas.
More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon
MongoMation is an autonomous PLC engineer. You describe a machine in plain English, and Claude agents write IEC 61131-3 PLC code, compile and deploy it to a…

LabRat is a self-improving harness for a physical lab agent, running on a real bench: two SO-101 arms, a 3018 CNC carrying a servo pipette I designed during…

Recursive Harness Scientist: an AI scientist that improves its own scientific method, and must replicate every gain on data it has never seen. Why it…

Super Bid is a self-improving AI estimator that teaches itself to price bids the way a specific company would. It practices on the company’s real past bids…

A robot that gets better at handling objects it’s never seen, without retraining. After every round, an AI coach reviews what broke and rewrites the robot’s…
WORLDSTATE gives physical spaces a memory. Any phone becomes an observer: open a link, point the camera, and WORLDSTATE tracks what's inside fridges, shelves…
