I Built 41 Repos in Five Months. The Green Checkmarks Were Lying.

But the number worth asking me about is a different one: zero. That's how many times adapter conformance had passed when I audited it on July 1, behind two months of green checkmarks. The gap between those two numbers taught me more than the 41 did.

March in Omaha is good building weather

It's gray, it's 34 degrees and there's nothing else to do.

January was nine commits of poking at an idea. February was scaffolding. March was 358 commits across 24 repositories, 22 of which didn't exist on the 1st. By April 9, all eight pipeline layers were built. 580-some GitHub contributions for the year, nearly all of them inside that window. I've been shipping software for 30 years and I have never produced anything close to that rate. Nobody has, by hand.

The honest mechanics: I barely wrote code. I wrote rules, templates, prompts and validators, then let the model generate against them, with every artifact frozen before the next one built on top of it. So when people ask whether the pace everyone brags about with AI-assisted development is real, I can answer from the far side of it. It's real. That was never the question for me.

The question was the one I'd spent three years asking other teams: how do you know any of it works?

Two months of green, zero real passes

On July 1, I went to verify a single adapter that was supposedly finished, and something in its pipeline caught my eye. The signing step was skipped. Not red. Skipped.

Pull that thread across all 13 adapters and the picture comes apart. Every conformance job ran under continue-on-error: true, a setting that forces the step to report success no matter what actually happened. The signing steps were gated on a genuine pass, and they were skipped in every adapter's latest run. A gate like that only skips when its condition is false. Conformance hadn't passed. Not once. On any adapter. Ever. The dogfooding loop I'd been describing in my own docs did not exist, and nothing was ever going to tell me.

The root causes were almost insultingly ordinary. One adapter's CI job never installed pytest, so the adapter had nothing to run and failed all four criteria. The other eleven read input keys the test harness never sent them. KeyError, before doing any work, on every run, since late April.

Twelve days after that audit I published . I wrote it about other people's agents. It's true of mine too, and I built the platform side specifically so that 80 wouldn't be improvised.

When people ask what five months of full-speed AI-assisted building is like, this is the answer I give now. Generation is the cheap part, cheap enough that the repo count stops being interesting. A green checkmark is a claim somebody configured. Evidence is what survives an attempt to falsify it, and the distance between those two was one unreviewed YAML key. And a governance system you won't run on itself is a slide deck. Mine caught its own author four times in one July (masked conformance, a phantom driver, an unenforced invariant, an encoding bug), and each catch was embarrassing for about an hour and then permanently load-bearing.

The stale title treated the count as the accomplishment. Five months on, the count is the least interesting artifact in the org. What I'd actually show you is the July 1 audit that called my own checkmarks liars, and the pipeline changes that made lying structurally hard afterward. AI will happily build you 41 repositories. Whether you can trust them is still your job.

Todd Linnertz is a Senior Solutions Engineer with thirty years of enterprise engineering experience. He's the creator of AIEOS, an open-source AI governance system for software delivery teams. Find him at .