Developers Over-Engineered AI Agents but Left Their Own Workflow Unstructured

A software engineering analysis argues that developers have built extensive tooling to keep AI coding agents on track — including linters, CI pipelines, and rules files — but have applied almost no equivalent structure to managing their own work across multiple tasks. The piece distinguishes two distinct loops: an execution loop focused on completing a single task correctly, and an orientation loop concerned with maintaining a shared, accurate picture of all ongoing work. Most existing tools address only the execution loop, leaving the orientation loop to informal methods like manually updated markdown files. Research from ETH Zurich, cited in the article, found that AGENTS.md context files offer minimal benefit to agents and can even slightly reduce performance, while costing over 20% more in context overhead. The author calls for 'harness engineering' to be turned inward, applying the same disciplined, automated accountability to the human operator's workflow that is already standard practice for AI agents.