A new framework teaches AI agents to think like a kitchen brigade: settle the division of labor first, then handle the chopping and plating separately.
Researchers published OverForge, a training-free hierarchical architecture for cooperative AI agents, on arXiv this week. Instead of mapping observations straight to actions like most multi-agent LLM setups do, it splits reasoning into two layers: a strategic layer that decides roles and division of labor, and a tactical layer that handles moment-to-moment actions inside each agent's own partner-conditioned model of the world. A module the researchers call the "Prefrontal Cortex Module" bridges the two, branching out strategy-action combinations, simulating their outcomes with a forward model, and committing once it is confident. In the OvercookedV2 cooking simulator, OverForge-equipped agents plated 7 soups in a shared kitchen, more than double the 3 managed by standard flat LLM agents, while sticking to agreed roles and adapting when an unfamiliar partner proposed a different one.
The real finding is not the soup count. It is that separating "what role am I playing" from "what do I do right now" lets agents keep coordinating even when a partner behaves unpredictably - a problem that trips up most current multi-agent LLM systems, which tend to improvise tactics without any persistent sense of who is doing what. The team's ablations and memory-restart tests back this up: strategies that persist across episodes also sharpen how well agents predict their partners' next moves.
OvercookedV2 is a toy kitchen, not a boardroom or a codebase, so whether this strategy-tactics split survives contact with messier real-world teamwork is still an open question.