Builder, directing AI agents
I build software with teams of AI agents, run by a workflow system I designed.
I decide what gets built and what counts as done. The agents write and test the code.
It built Crawft and two apps now on the App Store, and it built a full SAT course in about 72 hours.
An AI coding agent is a program that can write code, run it and check the result. One agent is useful. Ten working at once tend to collide: they edit the same files, skip the tests and burn through usage limits.
So I designed a system that runs them like a team. I describe what I want in plain English and decide what counts as done. The agents plan, build and test in parallel, and nothing comes back to me as finished without proof that it works.
Every request starts at the front desk, the one AI I talk to. Its job is to delegate, so it stays free to answer me while the work runs. Before it acts, Jev, a small routing model I set up, reads the request and what's already running, then judges how big, risky and splittable the work is: answer it now, hand it off, or plan first.
The front desk turns the work into briefs. Each brief names the outcome, the files that agent owns and the one check that proves it's done, and goes to the cheapest model that can do it well: a fast model for mechanical work, a standard one for most tasks, the strongest for hard judgment, or a second AI company's model for an independent build. Workers run in parallel, each in its own copy of the code, and a worker can never start more workers.
When a worker finishes, its evidence is checked against what I asked for. Payments, logins, deletes and deploys also go to a separate reviewer agent that can send the work back. Only then does it ship, and I get a short report.
How a request moves through my agent system
Most of my design work went into rules, because agents are fast, confident and sometimes wrong. They test every feature the way a user would, in a real browser or on a real device. Every task I give goes into a ledger, and no task can be closed without evidence of what was checked and what was seen. If the front desk starts grinding through work itself, a hook nudges it to hand off, so I can always reach it.
Jev also guards my attention. When the front desk wants to ask me something, Jev scores each question against a rubric I wrote: would my answer change the result, is it my kind of decision, could the agent look it up instead. At most three questions reach me, each with a suggested default. Before I trusted any of it, the system ran 60 end-to-end test sessions in a sandbox.
Making sure the code runs without errors is not ever enough.
A live dashboard on my Mac shows every chat and every agent as a small pixel-art room. This one is from the session that built this website: me at the top left, the main agent beside me, and five workers below, each on its own task, from curating photos to screenshotting finished pages.
My part during the day is direction and review. I check in between classes, answer what's waiting on me and read the reports. Anything only I can provide, like a login code or a call on design, lands in a Needs you list instead of stalling the work, so the agents keep going while I'm away.
Every product on this site came out of this way of working. These five show it best, from a full company product to an app built in a week.
The largest thing I've built: a Mac app that records each student's school day, a backend on AWS, AI that writes a daily report on each student, and the dashboards teachers read. Agents have built all of it from my specs since May 2026, and it's in beta on Macs at Alpha, my school.
Shown with demo data.
On September 29, four days before my own SAT retake, I asked for a course on the hardest parts of the test, built on learning science. Six lanes of agents worked at once. About 72 hours later it was live, with more than 12,000 practice questions that agents wrote and other agents checked, the skills behind each hard question type mapped as a graph that tracks what a student has mastered and brings skills back for review, a short video for every skill and a checkout page.
At Wharton's summer startup program I led a team of five and built NestIQ's iPhone app by directing agents. It gives parents of children 0 to 3 a short daily plan of screen-free activities, and it has been on the App Store since July 25, 2026.
TaskGo! lives in the Mac menu bar. Start a task and a floating timer counts down your estimate and sounds an alarm when time's up, and Done moves you straight to the next one. Every minute on a task earns XP toward 100 levels, a planner turns the week into calendar blocks, and Goals tracks bigger projects by milestone. By March an early version had an Activity tab charting my keyboard and mouse use, two months before Crawft took the same question to whole classrooms: where does the time actually go?
Ippo turns a run into a game. Mid-run the watch buzzes, a creature appears and you sprint to catch it, and the coins and XP you earn go to raising the pets you've caught. It has been on the App Store since February 2026, and I pitched it on a show filmed in the ocean off Oahu.
Some builds are experiments. At Stasis, an invitation-only hardware hackathon in Austin run by Hack Club, a nonprofit for teens who code, I built Drum Cockpit, a nine-pad drum kit that keeps AI voice dictation running while you drum and cuts it off when you stop. For School Game I scanned my school with an iPhone's LiDAR sensor and rebuilt it as a 3D replica you can walk through. My first project on GitHub, Maze Antidote, a pixel-art horror maze game, won my school's Horror Game contest.
Drum Cockpit, built at Stasis in May 2026.