Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.14767.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

A paired experiment (43 pairs, 86 runs) on a business-intelligence reporting task holds five LLM agents, models, prompts, tools, and data constant while toggling manager loop-back authority. Flat organization scores higher on Utility and Writing Clarity; hierarchical reports hedge 53% more, cost 51.5% more tokens, and the quality gap opens inside revision loops.

Isolates one variable most framework comparisons leave implicit. Yields a concrete design rule for open-ended multi-agent work: a supervisor is only worth its token cost when it can verify, not just critique.

Cited by

No citations on record.