A new paper argues the manager in multi-model AI systems does not need to be a model at all.
UnitBoost replaces the generative meta-agent (the LLM call that reads workers' outputs, writes the final answer, and decides when to stop) with a fixed operator instead. A task-specific unit map turns each worker's output into slot-value proposals, a constrained argmax assembles the final answer, and any slot left unfilled or unsupported becomes an explicit residual to chase in the next round. On three held-out benchmarks, it beat the best single candidate chosen with gold labels by 0.060 to 0.195 points, and beat input-matched generative managers by 0.048 to 0.076 points. Swapping in just this management step improved six different compound-system setups by 0.013 to 0.182 points, and directing extra rounds at the flagged residual pushed FanOutQA cell F1 from 0.4778 to 0.5524.
That matters because compound LLM systems, the multi-agent setups increasingly common in agentic tools, usually hand coordination to another opaque LLM call whose output can shift depending on the order results arrive in. UnitBoost's operator is order-free and records exactly which worker's output fed which part of the final answer, turning a black-box judgment call into something you can audit and debug.
It is not a universal fix. The paper's own tests show the gains vanish when there is only one output to choose from, when you cannot tell which unit produced which result, or when every extra check has a real cost - a useful reminder that a merge rule beats a manager only when the problem is actually decomposable.