Coding agents tune a warehouse robot planner, with trade-offs exposed by simulation

Researchers used coding agents to explore trade-offs in a warehouse robot planner.

An experiment called Late Autumn asks whether coding agents can adapt a general multi-robot planner to one warehouse. In the November 2 preprint, the researchers ran six campaigns using a simulated floor modeled on an existing automated warehouse. The planner combines job assignment and route planning. The agents worked on the planner’s code, while the simulation measured how its changes affected throughput. [Late Autumn preprint](https://arxiv.org/abs/2611.02391)

More throughput came with conditions

The authors report throughput improvements of 22% to 60% in the evaluated campaigns, with gains on previously unseen random seeds. Some improvements transferred to another layout, a larger fleet or different service times. They shrank with a smaller fleet and mostly disappeared when the workload was heavily weighted toward outbound tasks. Some patches increased planner computation by more than 100 times.

Those results illustrate the difference between improving a simulation score and improving the operating envelope of a system. A planner may move more work during a run while needing too much computation to meet a real decision deadline. A change that helps under typical demand can lose its advantage when the mix of work changes. [Late Autumn preprint](https://arxiv.org/abs/2611.02391)

The code still needs review

The paper says final changes consisted of chains of scored commits, with a median step of 27 lines. That gives reviewers a way to isolate individual changes. The authors found problems in review that the throughput score had missed. A practical next test would include a compute budget, unusual demand, degraded equipment and the ability to revert a change. These are separate from whether a patch improves the average score.

The research concerns coding agents tuning planner software offline. It does not show an AI agent controlling a live fleet. Its value is the testable comparison between general-purpose planning and site-specific optimization, along with the evidence about when the gains weaken. [Late Autumn preprint](https://arxiv.org/abs/2611.02391)