DEVOURED
Why Software Factories Fail
The 'lights-off' software factory is failing because LLMs cannot maintain codebase quality without human architectural oversight.
What: Dex, founder of HumanLayer, argues that automated agentic workflows suffer from a lack of long-term maintainability, as models lack the fast-feedback oracle required to reward good architectural design during training. He recommends a return to a four-phase human-steered process: product review, system architecture, program design, and vertical slicing.
Why it matters: The industry is experiencing a 'velocity trap' where companies equate high PR volume with actual progress, ignoring the long-term technical debt generated by agents that optimize for code completion rather than system sustainability.
Takeaway: Adopt 'vertical slicing'—designing and testing a feature as a complete functional slice (e.g., API to UI) rather than layer-by-layer—to catch architectural flaws early in agentic workflows.
Deep dive
- Lights-off software factories (zero human reading/writing) are currently failing in production settings.
- Models are RL-trained to maximize pass/fail test results but not for code maintainability or design quality.
- Current benchmarks (like SWE-bench) are insufficient for evaluating codebase quality over time.
- Humans must re-introduce architecture and program design steps to maintain quality.
- Agent output quality degrades significantly after 3-6 months in 'agent-native' codebases.
- Vertical slicing allows developers to 'resteer' agent output before code debt accumulates.
- Reviewing 100-200 lines of code at each slice is more efficient than reviewing large, model-generated pull requests.
Decoder
- Lights-off software factory: An automated development pipeline where AI agents handle coding, testing, and deployment without human intervention or code review.
- Shotgun surgery: A maintenance anti-pattern where a single change requires modifications to many different files or locations, often caused by poor initial design.
- Vertical slicing: A development strategy that builds a functional feature end-to-end (database to frontend) in small increments, rather than building complete layers (database, then service, then API, then UI) separately.
Original article
Full article content is not available for inline reading.
Explore all 60 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi