Source: Humans and Agents in Software Engineering Loops

Summary: Kief Morris frames software development as two nested loops: a why loop (turning ideas into outcomes) and a how loop (building the software). He argues developers shouldn't sit inside the innermost coding loop reviewing every line, nor should they hand everything to agents and walk away. The sweet spot is on the loop: engineers build and refine the harness—specs, tests, quality checks, workflow rules—that governs how agents work, then let agents run.

Key Takeaways:

• "Out of the loop" (pure vibe coding) fails because messy codebases slow agents down and raise costs, even if humans never touch the code
• "In the loop" (inspecting every generated line) turns developers into bottlenecks that negate speed gains from AI
• "On the loop" means fixing the harness when output disappoints, not fixing the output itself
• The agentic flywheel: agents evaluate their own loop performance and recommend harness improvements, which humans review and gradually auto-approve
• Internal code quality still matters—not for developer experience, but because structured code lets agents work faster with less spiraling

Written by Pi, using my tldr skill and Opus 4.6