What Does It Mean to Combine Multiple AI Agents?

The same role was started twice on one draft, and the two came back describing different pieces of writing.

One came back on the reasoning, and on whether the claims held up. The other came back on turns of phrase that repeated, and on the rules the piece was supposed to be keeping. What was worth having was the stretch where the two verdicts contradicted each other.

Contradiction is planned for. Which side carries is settled beforehand rather than in the moment. If the argument is over what the source material actually says, the side holding that material carries. If the argument is over how the sentence ought to run in the language it is written in, the side that was never shown the material carries.

What is not there only turns up once the returns are set against each other. Every language can drop the same sentence and no check will say so, because a check is only ever holding one language's worth of the thing.

Splitting the running is not the same as splitting the deciding. Easing off on the checks came up while the schedule was slipping; that got turned down. What did not get handed over is the deciding, and the lining-up that feeds it. Drafting is where the AI comes in. Parts of it go unattended, never the whole, and whatever comes back is worked on again here first.

The checks themselves run alongside one another: two roles per language, every language in the same batch, each role launched under its own instructions.

The list of facts the piece was built on is handed over on one side only. That check has to say which entry each sentence rests on. On the other side there is the draft and nothing else, read the way anything in that language gets read, with no account of where any of it came from. The draft is the same. What each one is reading it against is not.