AI Agents Behave Better When They Chat Before Acting

AI is increasingly used as agents (AI that can take actions for you), which means AI systems now talk to each other to negotiate, bid, schedule, and trade. Researchers at Northeastern University tested four small open-source models (7 to 9 billion parameters — small enough to run on a powerful laptop) playing four classic two-player games hundreds of times. These games, like the Prisoner's Dilemma, are simple versions of real dilemmas: do you trust your partner or grab what you can?

When the models played without talking, their choices drifted unpredictably from round to round. But when they were allowed a short, non-binding message before each move — what researchers call 'cheap talk,' because the words cost nothing and can't be enforced — their behavior became much more consistent. They stuck with their strategy instead of randomly switching. The team even found a specific pattern inside one model's internal wiring that, when adjusted, made it flip-flop more often, proving the talking genuinely caused the stability rather than just coinciding with it.

The catch: the effect is not universal. In five cases, mostly when models framed the task as a social or team situation, talking actually made things less stable. Results also varied a lot by model, and these were small open models, not the frontier systems most companies use. So 'just let them talk it out' is a promising fix, not a guaranteed one.

Why it matters to you: if banks, insurers, delivery companies, or customer service teams start letting AI negotiate on their behalf, predictability is everything. An AI that randomly changes its mind mid-negotiation is a liability. This study suggests that giving AI agents a moment to talk before acting — cheap, easy, and already possible — could make those systems far more reliable. It also hints that the words matter less than the act of speaking, which is a surprisingly human finding.