An agent's "done" is a claim. The check is the proof
An agent says "done". What do you know then? That it thinks it is done. Nothing more.
The more room you give agents, the less you can lean on what they say themselves. That is why every ticket in my setup has a check. An agent may only report "done" when the check confirms it.
An example from the server
When ordering the server (part 1), the claim was: "the server runs". The check was: can you log in, and does NonStopWorker answer? I do not need to believe the agent. I let the check run.
NonStopWorker
NonStopWorker is the controller I use to keep agents working, even when one of the AI engines drops out. It takes a ticket, lets the agent work and runs the check. Only when the check passes does the ticket count as done. If the check fails, the ticket goes back to the agent with a note on what was wrong.
A check has to be proven itself
A check that is always green proves nothing. So I break what it guards in a copy and see whether it turns red. I wrote about that earlier, in the piece about the green light and the report.
What a check is not
A check is not a second opinion from the same agent. An agent checking its own work with the same assumptions will not find the same mistakes. A good check stands apart from the work: a test, a measurement, a call from outside.
What you can take from this
- Write down per task what proves it is finished, before the agent starts.
- Let something other than the agent run the check.
- Make sure "done" cannot happen without a passing check.
- Break the check once to see it turn red.
Which task in your organisation is "done" now because someone says so, and what would the proof be? I am happy to work that out with you in an Expert Day.