The current software factory debate is not really about whether coding agents are useful. The sharper question is whether teams can remove human review from production development by adding more harnesses, sandboxes, automated checks, and agentic loops. A HumanLayer essay argued that this promise is much weaker than it sounds, and Hacker News treated it as a live argument rather than another anti-AI broadside.

The essay starts with the industry narrative: developers are the bottleneck, models are good enough, code is cheap, and teams should ship more. It references lights-off software factory experiments, OpenAI's harness engineering framing, and the broader push to route tickets, incidents, and user feedback into agent workflows. In that model, the agent builds the change, other agents review and test it, and humans move further away from line-by-line code reading.

The counterargument is practical. Agentic coding can compress build time from days to hours or minutes, but review and validation do not disappear. In complex brownfield systems, the article argues, the real failure mode appears after months of accumulated agent-built changes. A difficult issue eventually forces engineers back into a codebase they have stopped reading, and the saved review time becomes a debt.

The author also rejects the familiar “skill issue” explanation. Better prompts and larger token budgets can help, but the post argues that no amount of harness engineering fully solves what is partly a model training and evaluation problem. Benchmarks can reward local task completion while missing long-term maintainability, product context, and operational risk.

The constructive part is old-fashioned: front-load alignment. Planning, architecture review, clear acceptance criteria, and human judgment still reduce rework. The community discussion focused on that bottleneck. The appeal of software factories is obvious, but the claim that human review can be erased remains the part that needs evidence.

HN discussion / source

#software-factory

Why the Software Factory Debate Is Really About Review, Not More Harnesses