Agent triage

Turn a to-do list into a small set of well-specified agent jobs, get them approved, then run them.

The idea this rests on

Agents do not take tasks off a list. They change what state a task is in when the human arrives at it.

That reframing matters because of a trap that shows up immediately in any real list: the tasks worth the most money are the ones an agent must not touch. Choosing which product to back, settling a revenue split, deciding whether to end an engagement — those are irreducibly human. Meanwhile the tasks an agent can finish outright are usually the cheapest ones on the board.

If you only look for tasks an agent can complete, you will find the admin and miss the leverage. The real prize is agents doing the preparation for the expensive decisions — the research, the draft, the options memo — so the human's judgement lands on a fully prepared surface instead of a blank page.

So the rubric has four dispositions, not two. See references/triage-rubric.md for the full test and worked examples; the short version: