. @dexhorthy opens "Harness Engineering is not Enough: Why Software Factories Fail" with a bold claim about agentic coding: no amount of harness engineering or loops can fix what is fundamentally a model training issue. It's on @aiDotEngineer's YouTube. Dex is Co-Founder at HumanLayer. The talk gives you a mechanical explanation for why coding agents erode your codebase over time, traced back to how the models are trained, plus the workflow HumanLayer uses instead. - They tried lights off and it broke. In July 2025 HumanLayer stopped reading the code. Months later they hit a bug the agent couldn't solve, and had to dig into a codebase nobody had read since. - Brownfield starts sooner than you think. Dex argues agents begin to struggle on a codebase after roughly three to six months, not after ten years. - The training loop can't see maintainability. SWE-bench-style rewards are binary: did the test pass, did you break anything else. Nothing in that scoring penalizes a needless try/catch or a cast that exists only to make a test go green. - The cost function of bad architecture is measured in months and years. That gap is why the reward signal can't propagate back to the coding episode that caused it. - Cloud Code's edge was training against its own harness. Same tools as ADER and CodeBuff. The difference was a model lab RL'ing the model inside the harness it shipped in. - A judge model has a ceiling. If the model knew what good code looked like, it would have written it the first time. - Newer benchmarks are reaching for this. SWE Marathon from Abundant AI, Deep SWE from Data Curve, and frontier code from Cognition, which penalizes tests that don't fail on the pre-patch code. - Plan up front so review stays cheap. Product review, system architecture, then program design (types, method signatures, call stacks), then vertical slices for order of implementation. - Program design is the underemphasized step. People assume the model can cook once architecture is right. Dex says the layout and call stacks are where the work is. - You don't have too many PRs, you have too many bad ones. A good PR is a joy to review. Thirty minutes of alignment up front saves hours of review, which makes reading every line still feasible. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!