Beyond Coding Why Coding Agents Keep Making Your Codebase Worse
Sep 30, 2026
Dexter Horthy, CEO of HumanLayer and the engineer who coined “context engineering,” examines why AI coding agents can quietly worsen codebases. He introduces SlopCodeBench, challenges SWE-bench hype, and explores dark factories, token maxing, and realistic 2–4x productivity. Horthy also discusses design docs, engineering taste, pull requests, janitor bots, overnight agents, and focused parallelism.
AI Snips
Chapters
Transcript
Episode notes
Why SlopCodeBench Exposes Long-Term Code Decay
- SlopCodeBench tests whether agents maintain a codebase across successive feature additions, unlike narrow SWE-bench tasks.
- Models can pass isolated tests while steadily increasing failures and complexity over time.
Maintainable Software Remains The Hardest Frontier
- Maintainability remains a difficult frontier even as models improve on computer use and other capabilities.
- Dexter Horthy warns that the final 10% of software quality may be harder than the first 90%.
Plan Big Features Before Agents Write Code
- Align on architecture before asking an agent to implement substantial work, but accept that plans cannot capture everything.
- Use progressive design discussions to reduce rework from 25–75% toward roughly 0–10%.