The original title is "AI coding agents: everyone harnesses the agent's loop. Here's the human's."

Short summary

The article distinguishes two loops in AI coding workflows: the execution loop (is this one task right?) which is well-tooled, and the orientation loop (do you and the agent share an honest picture across all tasks?) which is left to markdown files and hope. It maps tools on two axes—execution vs orientation, advisory vs harnessed—revealing an empty corner where orientation-loop tooling should be. ETH Zurich research shows advisory files like AGENTS.md barely help and cost 20% more context to run.

• •Identifies a gap: agent execution loops are well-tooled but human orientation loops rely on manual markdown files
• •Maps tools on two axes: execution/orientation and advisory/harnessed, revealing an empty quadrant
• •Cites ETH Zurich study showing AGENTS.md files marginally help or slightly hurt while costing 20% more context

Generated with AI, which can make mistakes.