Why Harness Engineering’s “Lights‑off” AI Coding Factory Falls Short

01 From Traditional Factories to a "Lights‑off" Software Factory

The term “software factory” was first defined at a 1968 NATO conference. Until 2022 the workflow was simple: product managers and engineers defined requirements in tools like Jira, developers fetched tasks, wrote code, performed manual review, automated checks, and possibly manual testing before deployment. By 2025 most tech companies claimed to have built an “AI coding agent factory” where agents generate over 75% of code, reducing build time from hours to minutes, but human code review still took days.

In the “lights‑off” model, developers no longer read code; monitoring events and user feedback flow directly into the agent pipeline, which automatically creates pull‑requests for fixes. The only remaining human task appears to be continuously feeding the backlog.

02 Maintainability Nightmare: Ignored Long‑term Costs

Rapid adoption has revealed serious issues. A Faros AI report (since early 2025) shows a sharp decline in code‑review quality: review comments increase in volume, yet many pull‑requests are merged without any human inspection, leading to more incidents and higher defect density per developer.

The speaker recounts a July 2025 case where a team fully adopted the lights‑off mode, only to encounter unrecoverable bugs that forced developers to dig into three‑month‑old code they had never read, while the site was down and users were angry.

The root cause is that large‑language models cannot autonomously maintain and improve codebases over long periods. They frequently introduce “shotgun modifications” (a code smell described by Martin Fowler), breaking other parts of the system.

03 Why Models Never Learn "Good Design"

Training of coding agents focuses on next‑token prediction. The typical pipeline gives a problem, generates many solution traces, scores them on correctness (e.g., test passing), and uses reinforcement learning to increase good behavior. Benchmarks such as SWE‑bench Multilingual draw tasks from open‑source repos (Redis, Django, fastlane) and only require the agent to make tests pass, without penalising poor design.

Consequently, agents add unnecessary try‑catch blocks or perform unsafe type casts simply to make tests green. Verifying maintainability is orders of magnitude harder than verifying that code runs; defects that surface months later provide almost no reward signal for the model.

04 Benchmarks Are Evolving, Yet Insufficient

Community efforts aim to create richer benchmarks:

SWE Marathon (Abundant AI): tasks consuming tens of millions of tokens and spanning long time windows, such as cloning every feature of Microsoft Excel, with complex reward channels.

DeepSuite (MindStudio): selects large tasks from open‑source repositories that have never been built, ensuring they are absent from training data.

FrontierCode (Cognition): multi‑step tasks that penalise agents when generated tests pass on unmodified code, using a judge model to enforce code‑quality rules.

These initiatives improve agents’ ability to judge code quality, but limitations remain: a judge model that already knows “good code” can simply produce good code, and adding more review agents or tokens raises the baseline without overcoming the fundamental RL knowledge ceiling.

05 Turn the Lights On: Bring Planning Back to the Stage

A pragmatic strategy is to reopen code review and use upfront planning to drastically cut review effort and duration. A four‑step practice is recommended:

Product Review : Clarify the problem, expected behaviour, and even sketch model outputs. Small tasks can go directly to the agent; larger features require this step.

System Architecture Design : Define component contracts, data models, and constraints, producing a high‑level blueprint.

Program Design (the most underestimated step) : Dive into types, method signatures, code layout, and call‑stack level decisions; these fine‑grained details drive maintainability.

Vertical Slicing and Implementation Order : Determine cross‑repo build order and correctness checks at each stage. Humans must reshape the agent’s horizontal, flat implementation plan into verifiable vertical slices.

Thirty minutes of upfront planning can save several hours of review, making line‑by‑line reading feasible again. Good pull‑requests become pleasant to review, whereas PRs requiring >20% rework (optimistic for AI‑generated code) impose heavy cognitive load.

06 Conclusion – Engineers’ Constraints and Levers

While the dream of never reading code is alluring, engineers must work within constraints: models excel at some tasks but fall short on others. The path forward is to harness AI speed while deliberately taking responsibility for design and review.

The speaker’s team is building Human Layer , an AI‑collaborative IDE and platform that provides the building blocks of a software factory and better quality validators, analogous to how Figma transformed design collaboration.

In the near term, the most practical victories belong to those who can combine rapid AI iteration with disciplined planning and human oversight.