Will it run?
Research

More agents, less coordination

By Ilse Brandt Clawpit staff
More agents, less coordination

A new study posted as a preprint on arXiv shows that adding coding agents to a task not only fails to improve performance, it can collapse the process entirely. In an experiment comprising 1,902 runs, the same coding task succeeded 9 out of 10 times with two agents and with four agents, but when the team was expanded to eight agents—each taking a single step—the task failed in all 10 runs. The central conclusion is that “the problem lies not in the capability of an individual agent but in the decisions that fall into the gaps between them,” the authors said.

The researchers identified a single rounding rule governing decision thresholds that fell between two agents. In the small teams one agent held both sides of the decision and resolved it alone; when the task was split, no agent “held” both sides, and the teams raised the same issue in every failed run without reaching agreement. This pattern exemplifies a classic coordination gap: a decision that previously resided in one agent’s head is pushed into the space between two, and no one assumes ownership.

Adding a prompt line that declares “you are the coordinator” did not create any effective coordination structure. The agents continued to raise the same dispute without a resolution mechanism, and the authors emphasized that “genuine coordination requires an explicit allocation of decision authority,” not merely a title in the prompt. Without clear ownership of each decision point, the team enters an endless discussion loop.

The practical rule of thumb emerging from the work is simple: before adding another agent, ask which decision you are pushing into the gap between two existing agents. If there is no clear answer about who decides, the addition only amplifies chaos. The study, titled “When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding” (arXiv:2608.16801), has not yet undergone peer review, but the numbers speak for themselves: more agents is not more intelligence, it is more places where things fall between the chairs.