Why Software Factories Fail (or: harness engineering is not enough)
github.com
July 23, 2026
31 min read
🔥🔥🔥🔥🔥
54/100
Summary
Advanced context engineering enhances the performance of coding agents by optimizing how they interact with and utilize contextual information. Effective implementation of this technique can improve collaboration between humans and AI in software development.
Key Takeaways
- Companies using AI coding tools have reported a decline in pull-request review quality, with more comments, longer comments, and numerous PRs merged without review.
- Incidents and bugs per developer have increased since the adoption of AI coding tools.
- The narrative that coding agents can eliminate the need for code review through harness engineering is challenged, suggesting that model-training issues remain unresolved.
- The concept of "just token harder" is critiqued as an oversimplification of the challenges faced in AI-assisted coding.
Community Sentiment
Positives
- Opus 4.5 was a massive step change in capability — it’s not perfect, but it’s significantly more helpful than previous versions.
- The idea of grounding LLM-driven implementation on normative specifications is a promising way to streamline the coding process without the headaches of messy code reviews.
Concerns
- Gating integration behind code review is futile; many engineers are already automating it, indicating that current systems are outdated and ineffective.
- LLMs currently can't do long-term planning or choose the right abstractions, which is critical for maintainability — this highlights a fundamental limitation in their design.
