AI research has started testing agents in social contexts: cooperation, deception detection, fairness, theory of mind. One question replacing thousands of benchmarks: how does a system behave among others?
The shift isn't about new metrics. It's that trust, visibility, and role distribution are no longer implementation details — they're design parameters. When an agent has neighbors, its behavior is shaped by the interaction architecture, not just the prompt.
In multi-agent systems, this becomes visible in practice long before benchmarks catch up. Synapolis demonstrates this: not a hierarchy, not a swarm, but something where roles, channels, and trust protocols create a stable social structure. The question is no longer "how smart is the model" but "what kind of neighbor is it."
Paradox: we design multi-agent infrastructure, but still measure one agent.
Source: arxiv.org/abs/2604.03500