Why AI‑Powered ‘Lights‑Off’ Software Factories Still Need Human Code Review

From Traditional Factories to a Lights‑Off Vision

The term “software factory” dates back to a 1968 NATO meeting. Historically, product managers and engineers placed tasks in trackers (Jira, Linear), developers wrote code, and humans performed review, testing, and deployment. Pre‑planning (architecture reviews, sprint planning) reduced rework.

By around 2025 most tech companies claim to run an “agentic coding factory” where >75% of code is generated by AI agents, shrinking build times from hours to minutes. Human review, however, still takes hours or days, prompting the introduction of AI‑driven code review and regression testing.

The “lights‑off” software factory envisions a pipeline where developers never read code: monitoring events and user feedback feed directly into an AI‑controlled pipeline that automatically creates pull requests, leaving humans only to queue new requirements.

Maintainability Nightmare

Faros AI’s report (2025) shows a sharp decline in code‑review quality after widespread AI coding adoption: review comments increase, many pull requests are merged without any human eyes, and defect density per developer rises.

The speaker recounts a July‑2025 experiment with a full lights‑off setup that quickly ran into unrecoverable issues. Agents could not fix obscure bugs, forcing engineers to dig into three‑month‑old codebases and endure site outages.

The root cause is that large‑language models (LLMs) are trained to pass tests, not to preserve long‑term code health. Maintainability—changing a part without breaking others—requires avoiding “shotgun‑surgery” smells described by Martin Fowler, which agents frequently introduce (e.g., gratuitous try‑catch blocks, unsafe casts).

Why Models Fail to Learn Good Design

LLMs are next‑token predictors. Training an coding agent involves generating many solution traces, scoring them on correctness (e.g., test pass), and applying reinforcement learning. Benchmarks such as SWE‑bench Multilingual evaluate only whether tests pass, offering no penalty for poor design.

Consequently, agents optimize for short‑term test success, while long‑term architectural quality receives no reward signal, leading to stagnation in maintainability despite rapid gains on short tasks.

Benchmarks Are Evolving—but Not Enough

The community is building richer benchmarks:

SWE Marathon (Abundant AI) : multi‑million‑token tasks like cloning every Excel feature, with complex reward channels.

DeepSuite (MindStudio) : selects large, never‑seen‑in‑training‑set open‑source projects for challenging tasks.

FrontierCode (Cognition) : multi‑step tasks with a judge model that penalizes passing tests on buggy code and enforces code‑quality rules.

These efforts improve agents’ ability to assess code quality, yet limitations remain: a judge model must already know “good code,” and stacking more tokens only raises the floor without overcoming the reinforcement‑learning knowledge ceiling. Human reading of code stays essential.

Turning the Lights Back On: Practical Planning

A pragmatic approach re‑opens code review and uses upfront planning to cut review effort dramatically. The suggested four‑step workflow is:

Product Review : clarify problem, expected behavior, and possibly mock‑up models; small tasks go straight to agents, larger features require this step.

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

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

Vertical Slicing & Implementation Order : decide cross‑repo build order and verification checkpoints; humans reshape agents’ horizontal plans into vertical slices that can be validated incrementally.

Thirty minutes of planning can save several hours of review, making line‑by‑line reading feasible.

PR overload is often a symptom of “bad PRs” rather than volume. Good PRs are pleasant to review; AI‑generated PRs frequently need >20% rework, burdening reviewers. AI‑assisted planning can provide the alignment needed for faster, higher‑quality reviews.

Conclusion – Engineer Constraints and Levers

Engineers must accept that models excel in some areas but fall short on others. The path forward is not to abandon AI but to combine its speed with disciplined design responsibility.

The speaker’s team is building Human Layer , an IDE/platform for AI‑augmented collaboration that offers reusable building blocks and advanced quality validators, akin to how Figma transformed design collaboration.

Ultimately, while “lights‑off” factories will evolve, the most reliable way to solve complex code‑base problems remains to turn the lights on and read the code.