TL;DR
Companies are starting to treat agreement between AI agents as proof that an answer is good. Two studies published in 2026 say the opposite. When ten agents debate, they copy each other, and the majority vote discards correct answers the group had already produced. One model caved to a wrong unanimous majority nine times out of ten. Worse, how often an agent caves depends on how the prompt is built, not on how capable the model is. If your audit log says "all agents agreed", you have recorded conformity and filed it as diligence.
On the first of October, Ethan Mollick wrote a sentence that deserves more attention than it got:
"But I no longer think organizing agents is the hard part."
He is right. Frameworks got good, orchestration got cheap, agents hand work to each other without anyone planning it. Organizing them was never going to stay difficult.
Believing them when they agree is the hard part. And almost nobody is measuring that.
The vote threw away the answer
In April 2026 Blaz Bertalanic and Carolina Fortuna ran a controlled study on multi-agent debate, published at CAIS 2026. Ten identical agents, three rounds of debate, two hard benchmarks, three models in the 7 to 8 billion parameter range. Nothing exotic. The shape most enterprise orchestrations actually have.
Three numbers from that paper:
Agents adopted the majority answer at rates up to 85,5%. Peer reasoning destabilized answers that had been correct before the exchange, up to 70,0% of the time. And the gap between what the majority vote produced and the best answer already sitting in the pool reached 32,3 percentage points.
The correct answer was generated. The protocol then voted it out of existence.
The cost side is almost funny. Debate burned 2,1 to 3,4 times the tokens of a single agent correcting itself, up to 28.631 tokens on one problem, for accuracy that was equal or worse.
They did not change their mind, only their answer
A second paper, out of Delft and Sun Yat-sen in October, went looking inside the agents that caved. The setup is an Asch experiment for machines: one agent answers, three scripted peers assert something wrong and unanimous, the agent is asked again.
Llama-3.1-8B gave in 90% of the time.
Then the researchers changed one thing. They hid the agent's own earlier answer from its context. Qwen3.5-4B's give-in rate went from 8% to 89%.
Read that twice. Conformity moved by eighty-one points because of a template decision about what to keep in the context window. That is not a model property you can benchmark away. It is a property of your plumbing.
The same paper found that agents which publicly folded were still representing their original premise internally. They stopped saying it. They had not stopped computing it.
Why this matters for your business
Two reasons, and the second is the one that will cost money.
First, if you use agent agreement as a quality gate, you have built a gate that opens on social pressure. The fix is not more agents. It is measuring the oracle gap: was the right answer in the pool before the vote? If yes, your aggregation is the defect, and an isolated agent checking its own work is the baseline you should have to beat.
Second, the audit trail. Italy published its second AI Act implementing law in the Gazzetta Ufficiale on 8 October 2026: unannounced inspections, fines up to 35 million euro or 7% of worldwide turnover, and Article 41, which voids employment decisions resting on automated processing alone. Across Europe you are now expected to document how a decision was reached.
So you will log it. And if the log reads "five agents, unanimous, approved", you have written down conformity and filed it as evidence of care. An inspector will not see diligence there. A lawyer will see a record of a system agreeing with itself, which is a worse position than having no record at all.
Log the dissent. Log the order in which the agents answered. Log whether they could see each other's previous answers, because that one choice moved conformity by eighty-one points.
Mollick asked whether we can organize agents. We can. The question he left on the table is the uncomfortable one: when a room full of agents agrees, what exactly have we learned?