Done Means Verified
"Done" is the most expensive word an agent can say without proof. The words sound the same whether the change works or fails. Treat "done" as a claim until a check runs the part that changed.
By "done" I mean shipped, checked, and yours to stand behind. A claim becomes a result when something independent looks and agrees. Independent means something ran the code.
Verification model
A kid says "I cleaned my room." Do you believe the words, or do you open the closet? "Done" is a sticker on the door until someone looks inside.
The agent can sound sure and still be wrong. Sounding right and being right are different skills. The room can be clean and the closet still full.
You ask for proof that touches the thing that changed. Show me it ran. Show me the test that hits the new part. Words are a claim. Checking is the result.
In my own work, this failure sank more agent runs than any limit of the model. The agent edits the code, and then it grades its own edit. A grader that also wrote the answer is a weak grader. Turn every "done" into proof before you believe it.
01The model is fluent, which is the problem
A language model is good at making things that look right. That is the gift and the trap in one. When it finishes a change, it explains in calm, sure words why the change works. The explanation comes from reading the change, and nothing ran. So "the agent says done" tells you almost nothing on its own. Checking starts there.
02What real verification looks like
An agent you trust and an agent you babysit differ in one thing: how each checks its own work. Checking has four steps. Run them in order, and start with the test that touches the new part.
- Run the tests on the part that changed. A suite can pass while the new part sits untouched. Watch the output of the test that runs it.
- Take the typecheck seriously. An error is a small alarm. Read it, even when it points somewhere you did not touch. That line is often where the bug hides.
- Make the old problem happen again. Run the exact thing that used to break and watch it work now. Anything short of that is hope.
- Stay a little suspicious. If the run looks too clean, look harder. Hunt for the case that breaks the change. Reasons it is right are cheap to collect and prove nothing.
03Verification is delegation's whole point
When one agent hands work to another, a single rule carries the handoff: check the work yourself. A boss agent that forwards the "done" it received adds one more sure voice and no proof.
The second agent earns its place by inspecting the work from its own angle. Remove that inspection and the chain turns into a relay of summaries. Every link owes you the same check.
Name the ending
This is the same rule as failing well. A named failure gives you something to act on, and a checked success does the same. Nobody can build on a sure "done" that went unchecked. You learn it was wrong after you already trusted it.
04The rubber stamp
Flip the diagram to the mode that skips every check. The agent makes the change, says "done," and runs nothing. From the outside it looks exactly like real success.
Nothing burns trust faster. The first few times the change happens to be fine, you start to believe the words. Then it fails on the one that mattered. An agent that says "done" with no proof handed you the closet. You open it later, when the fix costs more.
The cost
It took me a while to respect this. Checking costs time, and it can be faked. Tests are slow and sometimes moody. A typecheck passes on code that is still wrong. A check can be gamed: a test that exercises nothing, or a check chosen because it is easy.
Check the thing that breaks in front of a real person. Staying suspicious is the guard against a gamed check. It aims the test at the part you fear. A green check on the wrong thing sells you a calm you did not earn.
Decision
Never take "done" on its word: ask which test ran and what it showed. When your agent says it worked, can it open the closet and show you? Two things count: the test that ran the new part, and the old break that breaks no more. If nothing is attached, hold it and do not ship. Next comes the moment one agent starts making more of itself. Subagents.