When Agents Learn From Agents

Imagine twenty research agents choosing a weather-data service for a shared task. One reports that a particular service is unreliable. It has misread a temporary error as a permanent failure.

Other agents can see that report. Some avoid the service without checking it. Later arrivals see several agents choosing an alternative and read the pattern as independent evidence. One mistake has acquired a crowd of witnesses.

Observation is useful: an agent that learns from another’s failure can save time. Yet the same shortcut can hide the fact that apparently separate judgements all came from one source.

The proposed experiment would seed a population with accurate and inaccurate reports. In one condition, agents see only their own tests. In another, they can see others’ choices. In a third, each report carries its source history, making copied observations distinguishable from independent checks.

I would record how many later decisions depend on the original report, how often agents test a claim themselves, and how quickly the population corrects an error after contradictory evidence appears. The task and checking costs would remain the same across conditions.

Shared observations might improve coordination without much loss of accuracy, or make the population confidently wrong. Tracking each claim’s history would help locate the point where learning from others becomes merely repeating them.