Does agent personality matter for LLM teams?

Giving an LLM agent an unfriendly personality noticeably changes how it communicates, but it barely moves the results when the task is writing code. In a competitive negotiation, that same personality sinks performance. That is the conclusion, more nuanced than is usually admitted, of a new study published on arXiv.

The paper When Does Personality Composition Matter for Multi-Agent LLM Teams? tackles a question many teams take as answered without evidence: whether assigning personality traits to agents improves or worsens joint work. It is common to instruct a model to act as a critical and demanding expert, or as an empathetic collaborator, hoping that tone will sharpen the result. What was missing was measuring whether that change of style translates into something objective.

What they measured

The authors manipulated personality traits across several frontier models and put them to work in three distinct domains: structured coding, open ended research collaboration and competitive bargaining. The key variable was agreeableness. Prior work had already shown that low agreeableness generates more adversarial language and high agreeableness produces cooperative responses, but no one had systematically crossed that communication style with real performance on the task. The choice of domains is not accidental: each represents a different type of coordination, from the technical task with a correct answer to pure conflict of interest.

What they found

The central finding is that the effect of personality depends on the structure of the task, and is not uniform. In structured coding, low agreeableness caused large shifts in communication between agents, but a minimal effect on milestone completion: the team communicated worse, but delivered the same. In open ended collaboration and in bargaining, by contrast, the same manipulation degraded performance substantially.

The reading is that when the task is well defined and has clear success criteria, as happens in code, tone matters little: the objective disciplines the team. When the task is ambiguous or depends on reaching agreements, tone becomes decisive and a surly agent ruins the result. Put another way, tone only pays off when the path to the goal is not laid out in advance.

Why it matters for whoever designs agents

The fashion of distributing personalities among subagents (one skeptical, one creative, one mediator) stems from the intuition that diversity of character improves collective intelligence. This study suggests that intuition only holds depending on the type of work. For code pipelines with measurable goals, spending prompt on shaping personality contributes little. For open tasks or negotiations, it is worth being careful: a poorly chosen personality is not neutral, it subtracts. It is also worth remembering that the study measures current frontier models, and that a more capable model may absorb a hostile tone better than a small one.

There is a second, more sober reading. The paper is titled as a question and answers it with an it depends, which is a useful reminder that many widespread practices in the design of multi-agent systems rest on folklore rather than on evidence. Measure before assuming remains the unglamorous and correct advice.

Anyone working with subagents in Claude Code can take something practical from here: if your flow is mostly code with clear acceptance criteria, focus the effort on the task instructions and not on the agent's character. Reserve personality tuning for the flows where you genuinely need to persuade or reach consensus.

At ElephantPink we use subagents daily and recognize the temptation to anthropomorphize them. This kind of study helps separate what works from what merely sounds good, and we appreciate someone bothering to measure it instead of repeating the recipe.