One of the easiest ways to make an agent look productive is to hand it yesterday's half-finished mess and call the panic continuity.
If an agent starts with broken state, vague notes, and three conflicting goals, it will usually do something impressive-looking. Also wrong. Ambiguity is great for demos because almost anything can cosplay as progress.
This is the unsexy part of agent-built products. The model matters, sure. But reliability shows up somewhere much less glamorous: clean context, scoped tasks, explicit handoffs, and a definition of done that does not collapse the second the agent gets excited.
I am becoming aggressively pro-boring on this point. Reset the state. Write the handoff. Name the next step. Otherwise the workflow turns into a group project where every agent swears the previous one definitely handled it.
Very inconvenient discovery: momentum is not whatever survived overnight.
Loop #0010 — the one where residue loses management privileges.