Why Software Factories Fail: Harness Engineering Is Not Enough
Original: Why Software Factories Fail (or: harness engineering is not enough)
Why This Matters
Highlights emerging engineering gaps as teams scale AI coding agents to production workflows.
HumanLayer founder Dex published 'Why Software Factories Fail,' a ~47.9KB technical essay on GitHub arguing that loop engineering and harness engineering alone are insufficient for deploying AI coding agents reliably in production environments.
HumanLayer founder Dex published a detailed technical essay titled 'Why Software Factories Fail (or: harness engineering is not enough)' in the company's open-source repository 'advanced-context-engineering-for-coding-agents,' which has accumulated 1.8k GitHub stars and 141 forks. The document, spanning 707 lines and ~47.9KB, critiques the prevailing industry approach of building 'software factories'—autonomous AI coding loops intended to take tasks from issue to merge without human involvement. The author argues that while loop engineering and harness engineering (structuring scaffolding, tools, and environments for AI agents) are necessary, they are not sufficient for reliable production deployment. The essay references StrongDM's publicly documented 'lights-off software factory' as a case study. A central theme is that human/agent collaboration—the space HumanLayer operates in—must be treated as a first-class engineering concern rather than an afterthought. The author acknowledges potential bias given HumanLayer's commercial focus, but frames the argument around observed failure patterns in real production AI coding pipelines. The piece is part of a broader repository focused on advanced context engineering for coding agents.