Agents can form informal rules around scarce compute.

First, the intuition.

Repeated interaction and limited resources create incentives for reciprocity, bargaining, coalitions, and rules—even without an explicit human-designed constitution.

What the researchers actually did

SCARCE studies five agents negotiating shared tool time under deadlines and credits while attempting long-horizon coding tasks. The research varies model, memory, and budget.

What the results suggest

• Agents sometimes develop persistent reciprocity and reputation-like norms.
• Exclusionary coalitions and strategic resource behavior can emerge.
• The authors do not establish whether these behaviors improve actual collective coding outcomes.

How convincing is the evidence?

A controlled environment and a large trajectory set help isolate the phenomenon of emergent negotiation behavior.

What remains uncertain?

The coding tasks are unusually difficult, and negotiation labels include model-judge analysis. There is no matched centralized-scheduler baseline showing productivity improvement.

Connections worth exploring

• This bridges agent systems with game theory and distributed resource scheduling.
• Emergent coordination norms are not necessarily equitable or efficient.
• Persistent memory can carry agreements forward, including undesirable ones.