The Pulse
LLM Agents Lift Warehouse Planner Throughput by Up to 60%
A study by Yimin Tang and five co-authors reports that LLM coding agents increased a multi-robot warehouse planner’s throughput by 22% to 60%. The gains weakened under smaller fleets and mostly disappeared with an outbound-heavy mix of orde
AI.info Team · Written with AI · fact-checked ·
Six campaigns, one warehouse planner
LLM coding agents raised the throughput of a multi-robot warehouse planner by 22% to 60% in experiments reported in a paper submitted to arXiv on November 2, 2026. The authors tested six autonomous research campaigns, asking agents to adapt a general planner to the needs of one warehouse. They report that the gains also held on test seeds the campaigns had not seen.
The study, titled “Late Autumn: LLM Agents Specialize a General Multi-Robot Planner for One Warehouse,” examines a lifelong planner that combines task assignment with multi-agent path finding. Its simulated floor was modeled on an existing automated warehouse. The work focuses on a setting where a site’s layout and processes rarely change, while operators still use general planners across multiple locations.
Agents tuned for a site, not every warehouse
The authors frame the task as a trade: the planner gives up some generality in exchange for higher throughput at a particular site. They compare that approach to adapting a pretrained language model for a narrower task. Rather than having experts tune the planner separately for each warehouse, the campaigns let coding agents investigate and change the planner for the modeled floor.
The agents rediscovered several established mechanisms for multi-agent path finding, including corridor-exit priorities, local guidance and rollouts. Independent campaigns also identified the same bottleneck, according to the abstract. The reported gains largely carried over to another layout, a larger robot fleet and different service times, but the authors describe important limits to that transfer.
Smaller fleets and outbound orders cut the gains
Improvements shrank when the researchers used a smaller fleet. They mostly vanished under a demand mix weighted toward outbound orders. Those results show that the throughput increase depended on the operating conditions tested; the paper does not present the top figure as a result that applies equally to every warehouse configuration.
The reported patches ranged from 0.7k to 2.4k lines. Each consisted of scored commits, with a median step of 27 lines, a structure the authors say experts can review. That review mattered: it found problems the scoring measure missed.
Compute costs complicate the throughput result
Some patches multiplied planner compute by more than 100 times. The paper therefore reports a trade-off alongside the throughput gains: the adapted planner may deliver more output, but some versions require substantially more computation. The abstract does not give a single compute increase for all patches; it says some exceeded that 100-fold mark.
The authors argue that site-specific tuning can convert a warehouse’s particular constraints into added throughput at little engineering cost, while leaving patches available for expert review. They also point to a role for experts in designing metrics and fast simulators that reflect the real problem. In the experiments described, the strongest gains reached 60%, while outbound-heavy demand erased most of them.