Loop-Back Authority in LLM Agent Teams: A Paired Experiment on Flat and Hierarchical Coordination
This presentation examines a controlled experiment isolating one organizational primitive in multi-agent LLM systems: whether a Manager can reject a worker's output and require revision. Through 86 paired runs on business-intelligence reports, the study finds that granting loop-back authority does not improve quality and instead degrades strategic depth while increasing hedging by 53% and token cost by 51.5%. The mechanism appears to be compliance-driven revision rather than substantive analytical improvement, with implications for when hierarchical coordination helps or harms LLM agent teams.Script
Most multi-agent frameworks give managers the power to send work back for revision, but no one has actually measured what that authority does to output quality. The researchers designed a paired experiment that changes exactly one thing: whether the manager can reject a draft and demand a rewrite, or can only comment and route tasks without enforcement power.
When the manager can force revision, reports become more cautious. Hierarchical reports contain 53 percent more epistemic hedges like may or suggests, they run 38 words longer after revision, yet they score lower on strategic depth and conciseness. The flat architecture, where the manager can only advise, produces sharper and more decisive synthesis.
The revised drafts are not rewrites. The median revision retains 88 percent of the original as a common subsequence and preserves 85 percent of the vocabulary. Instead of reconsidering the analysis, the writer keeps the structure and simply softens its commitments, adding qualifications without changing the underlying argument.
The hierarchy does not buy better accuracy. Both architectures achieve near-perfect specification correctness because the critic already has a verification tool. What hierarchy does buy is a 51.5 percent increase in token cost and a 34 percent increase in runtime, without improving any quality measure and while degrading utility.
The damage is not inherent to hierarchy. It emerges when the manager exercises authority. In the 11 runs where the manager chose not to intervene, hierarchical reports are numerically indistinguishable from flat reports. Writing clarity declines monotonically with the number of revision loops, and the increase in hedging appears specifically after the manager demands a rewrite.
The boundary condition is verification. When the manager can check a concrete property, both architectures already perform at ceiling. When the manager can only express an opinion about strategic depth or conciseness, authoritative revision induces compliance hedging rather than substantive improvement. If you want to explore how organizational design affects language model coordination, or generate your own video explaining a paper, visit EmergentMind.com.
