Labeling agents as AI drops deal-making success by 15 points

Researchers tested AI agents negotiating with each other. Each agent knew, or did not know, that its counterpart was also an AI. Agents told the other party was an AI took 30% more rounds to reach a deal. They used 55% more words doing it. They succeeded 15 percentage points less often. The identity label changed how the agents behaved, not how capable they were. Designers often treat transparency about agent identity as a safe default. This research shows that telling one agent who it is talking to changes how it behaves, and the change is costly. The agents did not become more efficient. They became more stubborn. For a design team, the question is where agent identity gets disclosed and to whom. Showing a human reviewer that they are reading agent-produced work is a different choice from telling one agent that its counterpart is also an agent. Those two disclosures do different things. This study addresses only the second case, but it gives design teams a reason to test both.