The Other Seventy Percent

Twenty senior leaders in a conference room. Monday. An hour debating a single sentence about why our organization exists.

Good work. Important work. The kind of alignment that can’t be skipped or delegated. We walked out with a shared purpose statement and I believed in it.

Then I flew home and opened my laptop to the fourteenth fix of the week.

I’d been building all week on the Mentilead apps. Two Shopify B2B apps that share a common set of components between them, with an AI coding agent doing the actual construction. Here’s what a week of “AI writes the code” actually looked like.

Across both apps, the agent produced roughly 9,000 lines of new code, touched almost 200 files, and shipped six features. It also had to fix fourteen things that broke along the way. A piece of shared software used by both B2B Onboard and OrderFlow needed ten updates in seven days as each fix in one app revealed a new problem in the other.

The agent wrote all of it. Every line. And I spent the entire week directing, reviewing, and iterating.

Six features to fourteen fixes. That ratio is the real story.

Building the feature is maybe 30% of the work. The other 70% is what happens next. Problems that only appear when the app actually runs. A help button that looks perfect in testing but gets hidden behind a pop-up in the real app. A minimum order rule that works on one page but doesn’t kick in when customers upload an order from a spreadsheet.

Each of those ten updates to the shared components happened the same way. The agent builds something. I check it. Something breaks or looks wrong. I screenshot the problem, send it to a design review agent, get critique back, pass the critique to the coding agent. Wait for the fix and check again. Something else breaks. Always something else.

Slow loops. Lots of waiting. The kind of iteration that’s boring to do and hard to automate.

The biggest feature of the week was a complete redesign of how wholesale buyers experience the ordering process. It started because I looked at the existing page and thought: this isn’t good enough.

Not a metric. Not a user complaint. A feeling I couldn’t have put into a prompt. I asked one AI agent to research best practices for B2B wholesale buyer experiences. It produced a thorough report. I turned that report into instructions for the coding agent. Three agents chained together manually: my judgment that something was off, a research agent that investigated, a coding agent that built the fix.

The result gives approved wholesale buyers a streamlined reorder experience while new buyers see a registration-first flow. Figuring out which type of buyer is visiting, designing two distinct paths, threading that logic through the entire app. None of that came from the agent. The request was “make this better.” The architecture was mine.

Same pattern with minimum order quantities. My AI COO spotted the gap from a Reddit thread during community seeding. Wholesale merchants need the ability to set minimum orders. Table stakes for the industry. But translating “merchants need minimums” into a real feature meant identifying six different places in the app where a customer could place an order below the threshold: the main ordering page, the admin settings, the quick order form, spreadsheet uploads, saved order lists, and the checkout itself. Each one needed its own enforcement logic.

Twenty-nine files changed. All because a human knew which doors to lock.

Back in that conference room, we spent an hour on vision and purpose because twenty people need to agree on direction before they can act independently. Alignment is the tax you pay for coordination.

The Mentilead experiment has no coordination tax. One founder, one direction. The agents don’t need alignment. They don’t have competing priorities or personal agendas. They do what they’re told.

But they also don’t know when something isn’t good enough. They can’t look at an ordering page and feel that it’s too bare-bones. They won’t read a forum thread and instinctively map every feature mentioned against their own product. The accumulated understanding of how real merchants actually work, that lives in the founder. Not in the code.

At the offsite, purpose was the thing that lets twenty people make independent decisions that still converge. Without it, everyone optimizes locally and the whole drifts.

Does a company of one founder and AI agents need the same thing? Not purpose in the motivational sense. But something that functions like it. A persistent sense of what “good enough” means. An awareness of direction that shapes every small decision without being explicitly invoked.

The agents don’t have that. I do. And this week it cost me every evening after work.

Anthropic published a pattern this month that maps to what I was doing manually. They call it the “harness.” Instead of one agent doing everything, you split the work into three roles: a Planner that defines the task, a Builder that executes it, and an Evaluator that tests the result and sends critique back. They report 5 to 15 evaluation cycles per feature.

That’s my loop. Build, check, critique, fix, repeat. Except I was the one deciding when to evaluate, which reviewer to use, and whether the output had crossed the bar.

I think the harness could automate 50 to 60 percent of what I did this week. The iteration cycle on a single feature is solvable. But the cross-feature insight (realizing that enforcing minimums on the order page means spreadsheet uploads also need the same check) and the cross-app coordination (fixing shared components that affect two products simultaneously) still needed a human holding the full picture.

I’m going to try it. Build a harness for the next feature sprint and measure whether that 6:14 ratio changes.

Autonomy score this month: 42%. Down from 50% two weeks ago.

The drop is real. This was a building week and every structural decision required the founder. The agent wrote 9,000 lines of code. I made every decision about where those lines should go.

The experiment’s thesis asks whether the only irreplaceable role is holding direction and values. Three weeks in, I’d add one thing to that list: knowing when something isn’t good enough yet. That instinct drove every iteration cycle. The agent never once said “this needs more work.”

Maybe that’s what purpose is for.