Building Legends: Build Software with Grokbot
A case study in agentic product development. For three years I had a product spec that failed on every platform I tried. With Grokbot running the team and Cursor cloud agents as the engineers, it got built in about twelve weeks. The human lane is taste and judgment.
Transcript
In January of 2025, I posted an experiment. I was moderating a debate between Niccolo Machiavelli and Jim Collins. Good to Great versus The Prince. I was experimenting with AI. JIM COLLINS: While this may achieve short-term control, it fails the test of endurance. Organizations and societies thrive on trust, shared purpose, and the compounding effect of disciplined action, not fear or opportunism. NICCOLO MACHIAVELLI: Collins, your faith in disciplined action and trust is admirable, but misplaced if it ignores human frailty. Fear, when judiciously applied, ensures stability where trust alone falters. A leader who builds only on shared purpose courts disaster the moment fortunes shift, as the masses are loyal to themselves above all. Shall we proceed? That was an example of the dialectical reasoning, and the opinionated AI, that I wanted early on. It evolved into the idea behind Legends. What if you could bring advisors into the meeting? Different perspectives in tension, and the human chooses.
People think it's about having a 600 IQ that gives you the right answer. It's the picking of the trade-off. Choosing in the trade-off. You say a name, and they appear in the meeting. One seat, named Legends, joins after the host authorizes it. You say hey, plus the name, and they answer out loud. They can argue. Legend Studio is where you create and tune a legend. The prompt, the voice, the wake phrase, without a deploy. I had the product spec three years ago. I tried it multiple times, with Claude, with Cursor, with every platform I could find. It always ended in failure, until I handed the spec to Grokbot. Here's how I built it. Building Legends, by building software with Grokbot. A case study in agentic product development.
I didn't hire a dev team. I designed a small one, and I worked with a product manager. I was the only human. Grokbot ran the team. Product management. Certification with Zoom. Live QA. Cursor cloud agents were the engineers. They wrote nearly all the code. I set the direction, and I stayed close to every test. I kept the hard-to-reverse calls. The product manager ran the day to day with the agents. Here's a concrete loop from early in the build. We called it a Ralph loop. You keep a task list and a definition of done. An orchestrator picks the next unfinished item. A specialist agent runs it against the spec, lands a small pull request, ticks the box, and goes back for the next one. It keeps going until the slice is done, or it's blocked and a human has to step in. Early on, done meant the whole path. Save the work, join the meeting, summon a name, and hear it speak. The loop just kept chewing through that list. A week in, on August 25th, we stood up CI. Every pull request had to typecheck and pass tests. Deploys stayed manual. One specific commit, after the release notes were written. If the tests passed and we could roll it back, we shipped it. Later on, if a change was easy to reverse and the tests passed,
we merged it and deployed it, including to Production, without waiting on me. For the heavier thinking, I used what I call the Big Brain Trust. Grok Heavy, Claude, or ChatGPT. I'd have them validate a plan, or review the code, come back with a punch list, and then the Cursor agents would clean it up. Testing meant live Zoom rooms. A QA bot would start a meeting, admit the seat, say a name or type it, and check that it joined and actually spoke. Agents will write a lot of code if you let them. Once we reached MVP, we cut it down. About 60,000 lines, down to about 15,000. I kept thinking, at some point a human needs to read the code. Then it occurred to me... who cares, as long as the AI can. About twelve weeks, from the first commit on August 18th. A hundred and seventy-nine pull requests opened, and eighty of them merged. I submitted the app to Zoom at the end of August. The review ran through September and October, and Zoom approved Production. It wasn't clean the whole way. When answers ran long, about 30 seconds, we tightened them the same morning, and a live test came back at 12 to 17 seconds. You check it, you fix it, and you go again. Everyone talks about AI slop. In my 30 years in the business, human slop is far worse.
I've had multiple engagements where the human-developed code was so bad it was cheaper and less risky to rebuild. That's my philosophy around AI. The agents don't have agency. I made the choices. The team, the loops, what to cut, what to ship. The human lane is making lots of choices around taste and judgment. AI can amplify what one person can build. The judgment and the vision... those still have to be yours.