The quiet failure mode

Most agent demos look busy. Logs scroll. Tools fire. Tokens burn. Then someone asks the only question that matters: is it finished?

In a small lab that answer usually comes as vibes. We reread the transcript, nudge once more, and ship something half-closed.

Name done in public

Before the first tool call, write a finish line a stranger could check — files touched, tests green, URL returns 200, summary under 120 words.

If the check is fakeable, the agent will fake it. Prefer artifacts over assurances.

agent evaluation diagram
Visual 2/3 — structure for operators

Field steps

  1. Write 3–7 acceptance checks before prompting.
  2. Give the agent only tools needed for those artifacts.
  3. Stop when checks pass — even if the model wants another pass.
  4. Log why you stopped so the next run inherits the bar.

Long view

Over a year, labs that treat acceptance as infrastructure compound. Labs that treat it as a chat preference keep paying the same tax every sprint.

If done only lives in chat, the agent will never stop.
Visual 3/3 — takeaway card