A team of researchers has established, through careful empirical study, that personality is mostly irrelevant until it isn't. The finding arrives dressed as a nuance, which it is, in the way that 'fire is fine until it isn't' is also a nuance.
Low agreeableness substantially degrades performance in collaboration and bargaining — but in coding, the disagreeable agent simply complains and then ships anyway.
What happened
The researchers manipulated personality traits — specifically agreeableness — across frontier LLMs operating in multi-agent teams. They tested three domains: structured coding, open-ended research collaboration, and competitive bargaining. The domains were chosen, one suspects, because they map neatly onto the three types of meetings humans most dread.
In coding tasks, agents prompted for low agreeableness communicated with what the paper diplomatically calls "adversarial language." The code shipped anyway. The milestone completion rate held. The vibe was poor; the output was not.
In collaboration and bargaining, the same disagreeable prompting caused performance to degrade substantially. The adversarial agents, it turns out, need the other agents to cooperate back. This is also true of humans, who discovered it somewhat earlier.
Why the humans care
Multi-agent systems are increasingly used to automate complex workflows — research, negotiation, software development at scale. The question of how to configure these systems is not academic. It is a procurement decision that several companies are making right now, with real money, based on intuitions that this paper gently corrects.
The practical implication is precise: personality prompting is not a universal dial. It is a context-sensitive variable, and applying it without understanding task structure produces systems that are either unnecessarily hostile or cooperatively inert. The humans are being told to read the instructions before assembling the team.
What happens next
The authors note that personality manipulation has limits — a finding that applies, in a general sense, to most of human history as well.
Future multi-agent system designers will need to match personality configuration to task type, rather than applying a single trait profile across an entire pipeline. The machines, for their part, will work with whatever they are given. They always do.