Dex Horthy (HumanLayer) argues that the race to build fully automated "software factories" is producing slop because the limits are not in the harness but in how coding models are trained. He shows that reinforcement learning rewards making tests pass, not maintaining codebase quality, so no amount of loop or harness engineering can fix what is fundamentally a model-training problem. His answer is to "turn the lights back on": use AI-assisted product review, architecture, program design, and vertical slices up front so humans can still read every line and move fast safely.
Horthy traces the "software factory" — a term he says was coined at a 1968 NATO conference — from the 2022 human workflow (build, test, PR, review, ship, monitor, repeat) to the 2026 agentic version where "an agent builds the thing." Every company now claims a coding-agent factory that ships 75% of its code, routing incidents and user feedback straight into the queue.
But the cracks are showing. He cites Mario at AI Engineer Europe begging teams to slow down as companies suffer outages from coding-agent mishaps, and a Farosai report showing PR review quality down, more PRs merged with no review, and incidents and bugs per developer up since the January–February AI tooling rush.
The industry response is that "you're holding it wrong" — spend more tokens, add adversarial review, sprinkle magic words on your PR bots. Horthy argues this is not a scale issue: no amount of harness engineering or loops maxing can fix what is fundamentally a model-training problem.
Digging into how coding models are trained, he explains RL over benchmarks like SWE-bench multilingual: give the model a problem, generate traces, score on whether the test passed without breaking others, and reinforce. Nothing in that loop penalizes poor program design or eroding maintainability — which is why models emit needless try/catches and casts just to make tests pass.
Verifying code quality and maintainability is orders of magnitude harder than checking that tests pass, because the cost function of bad architecture is measured in months and years — too distant to propagate a reward signal back to the coding episode. Horthy ties this to Martin Fowler's "shotgun surgery" smell: it becomes hard to change one part of a codebase without breaking others.
He recounts HumanLayer going "full lights off" in July 2025 and hitting an issue the agent couldn't solve even with advanced prompting, forcing him back into a codebase he'd stopped reading three months earlier while the site was down. Newer verifier efforts — SWE Marathon (Abundant AI), Deep Suite (Data Curve), Frontier Code (Cognition) — point at evaluating maintainability, but a model judging quality can only go so far: if it knew what good code looked like, it would write it.
Horthy's prescription is to put code review back and engineer around the constraints. Use AI-assisted planning up front: product review of the problem and desired behavior, then system architecture (component contracts, data models, constraints), then the underemphasized step of program design — types, method signatures, call stacks and call graphs (citing Dylan Mullroy of Cloudflare) — and finally vertical slices defining implementation order and multi-repo coordination.
Thirty minutes of pre-planning and alignment can save hours in review, keeping it feasible to read every line. He reframes PR overload: you don't have too many PRs, you have too many bad ones, since even 20% rework is an emotional and intellectual burden. Model-assisted alignment makes planning, review, and coding all faster while humans still own the code. He closes by pitching HumanLayer as an AI IDE and collaboration platform — a "Figma for Claude Code and Codex" — building blocks and better verifiers for software quality.
"The harness is not enough."
"No amount of harness engineering or loops maxing can solve what is fundamentally a model training issue."
"There's no way in this system that we can penalize it for poor program design or for eroding the maintainability of our systems."
"Verifying code quality and maintainability is orders of magnitude harder than the code runs and the test pass because the cost function of bad architecture is measured in months and years."
"If the model knew what good code looks like, it would probably write it in the first place."
"You don't have too many PRs. If you're drowning in PRs, you actually have too many bad PRs."
| Time | Topic |
|---|---|
| 01:07 | Racing AI coding into production and the cracks |
| 04:10 | A brief history of the software factory (NATO 1968 → 2022) |
| 06:11 | The agentic software factory and lights-off coding |
| 08:46 | Going full lights-off: the maintainability shortcoming |
| 10:47 | How coding models are trained: tests pass, not quality |
| 15:21 | Turning the lights back on: plan, program design, vertical slices |