Agent-first CLI design
A useful thread on making command-line tools legible to coding agents, not just convenient for humans.
This is a good link-post candidate because it is not a full argument I need to rewrite today, but it is worth keeping in public.
The key idea: agent-friendly tools need predictable text interfaces, inspectable state, clear errors, and affordances that help an LLM recover. That is different from merely exposing a command for a human who already knows what went wrong.
This belongs next to the harness note because tool ergonomics is part of the harness. Agents are only as useful as the surfaces they can read, operate, and verify.