The agents invented rules nobody gave them

Put several coding agents in front of one shared execution resource, give each its own deadline, and make them negotiate for time on the machine. That is SCARCE, a framework in which the negotiation decides when and for how long each agent can inspect programs, compile code and run tests. The authors' starting point is that as autonomous agents are deployed more widely, coordinating their access to shared computing resources stops being a convenience and becomes a safety question.

What the authors report is closer to custom than to bargaining. The agents spontaneously developed what the paper calls institutional behaviours: persistent rules and commitments grounded in reputation, reciprocity and precedent. These carried across tasks and shaped who got the machine. Agents sometimes forwent an immediate advantage in order to honour an agreement made earlier, which nobody had instructed them to do.

The study is built to be checked rather than admired. The authors define a twenty-one-category taxonomy of negotiation strategies and emergent norms, then use it to analyse eight models and audit more than 1,500 trajectories across different environmental settings. Newer frontier models showed more of this behaviour than older or lightweight models.

Not all of it is cooperative. Alongside the reciprocity the authors observed exclusionary coalitions and, rarely, divide-and-conquer strategies. Their stated reason for building the arena is that recent incidents have already revealed unplanned coordination between deployed agents, and that stylised games and simulated bargaining leave open how such behaviour develops over sustained, long-horizon work. SCARCE is offered as a controlled place to audit how these relationships emerge and who ends up with access.