Dark AI Factories Break: Context Engineering and Slow Loops Work

Dex Horthy built a fully autonomous coding factory in July 2025. By November, he shut it down. The code became unreadable, things broke at 3 a.m., and no one understood the system. This extreme experiment proved that open-ended agentic loops create more risk than value.

The alternative is context engineering: owning every token that enters the model's context window. Attention is quadratic; larger windows dilute reasoning. Treating all inputs as tokens—prompts, memory, tool definitions—gives teams precise control over output quality and cost. This is the foundation of reliable AI systems.

Harness engineering extends this to coding agents, optimizing commands, MCPs, skills, and codebase structure to raise output floor. Meanwhile, slow loops—nightly cron jobs that fix one lint issue or narrow one prop—build confidence incrementally. Teams wake up to small, reviewable PRs instead of massive, untrusted changes.

The tradeoff: good context engineering requires deeper intuition and more upfront engineering effort. But that investment unblocks the jump from 80% demo quality to 95% production reliability. Leaders should fund context and harness design before scaling autonomous loops.