← News

Agents & MCPSources: arXiv · 16 Nov 2026

OutfitterBench tests LLM agents on long-horizon resource allocation

A new benchmark asks language-model agents to run a simulated rental business for 200 days, where their own commitments shape future options. Most tested models trail a simple baseline.

In plain words

Researchers built a benchmark called OutfitterBench in which AI agents manage a simulated outdoor-equipment rental business. Agents must buy stock and handle reservations under a fixed budget, so early choices limit later ones. In tests, most language-model agents did worse than a simple rule-based program.

What was announced

The benchmark, introduced on arXiv, places agents in charge of an outdoor-equipment rental business spanning 200 simulated days. Agents purchase from a catalog of 46 products and handle roughly 1,100 reservation requests. Under a fixed capital ceiling, confirmed reservations and uncertain walk-in demand compete for reusable equipment whose availability shifts over time, so committing capacity creates future obligations and forecloses other opportunities.

In experiments with seven frontier LLMs, several agents underperformed a simple rule-based baseline, and only the leading agent consistently beat it. The authors' trajectory analysis found some agents applying known constraints inconsistently and making new decisions that conflicted with existing commitments. Agents that fulfilled orders reliably still sometimes earned low returns from comparable commitments or delayed reallocating resources. The leading agent connected preparation with economic choices better but also missed chances for more profitable asset use. The benchmark code is publicly available on GitHub.

What it does not show

  • The evidence is the paper's abstract, so experimental details, methods and full results are not available for verification.
  • The results come from the benchmark authors' own experiments and analysis; no independent evaluation is reported.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.