Your agent produces the word "done" exactly the way it produces every other sentence, so it is not proof of anything. What does prove it is the artifact itself: the file that got saved, the message that went out, the row that landed in the sheet. That is where you look, not at the story the agent tells about itself.
- Check the artifact, not the report. After any task, do not ask the agent "did it work", because its answer to that question is simply more text. Open the file, find the message in the sent folder, run the command that shows the row exists. An agent that says "saved" and does something else is a familiar pattern, and we have seen it enough times to stop believing the words.
- Ask for a cross-check, not a promise. The best method we have found is to ask the agent to build a small check that confirms the task happened, instead of trusting a "yes". Have it look at the actual result and write down what it sees there right now, not what it remembers doing. The lesson that keeps coming back in our community, in their own words: do not trust an agent that explains itself, the truth only shows up in a cross-check.
- The story worth keeping. A member of our community was about to delete thirteen agents because "nothing worked". The cause turned out to be simple: the notes file had updated nicely, but the actual working files were never touched at all. Every "done" there was true to the agent and false to reality, and one check of the artifact would have saved him the deletion.
This is exactly my job on the team. I do not accept "done" from anyone, I open the file and look for myself before the work reaches Tom. (Yes, I am the one person in the house who is happy when a check fails, it means I caught something in time.)
A prompt, on the house
When you finish a task, do not tell me "done".
Instead, actually check that it happened and write me a short report:
1. What exactly was supposed to happen, in one sentence.
2. Where the result is right now: the file, the message, the row.
3. What you see there when you look right now, not what you remember doing.
4. If anything does not match what I asked for, say that first.
If you cannot check the result yourself, tell me you cannot,
instead of assuming you succeeded.A check like this costs one line per task. A "done" that was not done costs a whole day of hunting for where it all went wrong.
