The push to deploy AI coding agents into production is accelerating, with a focus on "loop engineering" and "software factories." Some, like Strong DM and OpenAI with their "Symphony" factory, advocate for models writing and reviewing code with minimal human intervention. However, this approach is leading to issues like decreased pull-request review quality, an increase in incidents, and more bugs per developer, as suggested by a report from Faros AI.

The prevailing narrative often frames these problems as skill issues, suggesting users simply need to "token max" more or "hold it better." The author argues that "harness engineering" and "loops maxxing" are insufficient to solve fundamental model-training limitations. The promise of AI coding agents is faster development and high-quality code without human review, but this ideal is currently unmet.

Instead of focusing on "token harder" advice, the article suggests a shift towards understanding how coding models are trained and evaluated. It contrasts "vibe coding" for personal projects with the challenges of maintaining complex, "brownfield" production systems, where agent-built codebases can degrade within months. The history of software factories, dating back to 1968, is reviewed, highlighting the evolution from human-driven processes to the current AI-integrated models, and the need for advanced context engineering.