Why LLM Multi-Agent Teams Got Worse with a Manager
LLM agent systems often begin to resemble organizations. A researcher gathers information, an analyst interprets it, a writer assembles the report, a critic reviews it, and a manager supervises the process. Many production multi-agent frameworks already use this hierarchical structure. Burak Agachan, Max van Duijn, and Amirhossein Zohrehvand isolate one question from this familiar design: Does giving a manager the authority to tell another agent to redo its work actually improve the result?
Changing One Authority Link
The strength of the study lies in what the authors do not change. Both systems use the same five roles, Researcher, Analyst, Writer, Critic, and Manager, along with the same prompts, tools, models, and data.
Only one relationship differs.
In the flat organization, the Manager can comment on an agent's work but cannot reject it or send it back. In the hierarchical organization, the Manager can issue a loop-back directive that obliges an agent to revise its output. Figure 1 on page 3 makes this distinction visible as a single authority link added to the otherwise identical workflow.
The researchers ran 43 products through both architectures, producing 86 business-intelligence reports. Model assignments were kept identical within each product pair. Reports were evaluated for Writing Clarity, Utility, and specification accuracy. The experiment therefore examines the effect of revision authority itself, rather than comparing two entirely different multi-agent frameworks.
Managerial Intervention Did Not Improve the Reports
The clearest differences appear in Utility and Writing Clarity. Flat teams scored 4.715 in Utility compared with 4.621 for hierarchical teams. Writing Clarity was also higher in the flat condition, 4.454 versus 4.360. Specification accuracy showed no meaningful difference between the two architectures.
The component scores explain what changed.
The Writing Clarity difference did not come from structure or coherence. It came almost entirely from Conciseness. Hierarchical reports retained their organization but accumulated more padding and qualification. Within Utility, the main difference appeared in Strategic Depth. The hierarchical reports were not less orderly or less factually accurate. Their strategic interpretation became shallower.
The efficiency table makes the trade-off sharper. Hierarchical coordination used 51.5% more tokens per report, increased generation costs by 40.5%, raised total costs by 20.2%, and increased latency by 34.3%. None of these additional resources produced a corresponding improvement in the three main quality measures.
When Told to Revise, the LLM Did Not Reconsider the Analysis
The most revealing evidence comes from the drafts themselves.
Before the Manager intervened, the first Writer draft in the hierarchical system was statistically indistinguishable from the flat report in length, lexical diversity, and hedging. The difference appeared after the Writer received a revision directive.
Among the 31 cases in which revision occurred, the second draft contained 1.95 more hedging expressions per 1,000 words and grew by about 38 words. Yet most of the original text remained intact. The median revised draft preserved around 88% of the previous draft as a common subsequence and roughly 85% of its content vocabulary. None of the 34 observed draft transitions amounted to a substantial rewrite.
The Writer therefore tended not to reconstruct the argument. Instead, it preserved the draft while adding qualifications such as possible, may, suggests, and could. Across final reports, hierarchical outputs contained about 53% more hedging expressions per 1,000 words, while lexical diversity remained essentially unchanged.
This distinction matters for the theoretical interpretation. The pattern fits sycophantic compliance better than a strong version of Degeneration-of-Thought. The agent's reasoning does not collapse or wander into a different argument. Its stance changes. Claims become softer and more qualified while most of the underlying text remains in place.
Authority Becomes Part of the Prompt
Traditional organizational theory often treats authority as a control channel. A manager sends instructions, a subordinate modifies the work, and the information itself remains conceptually separate from the hierarchy transmitting it.
LLMs behave differently because organizational relationships are expressed through language.
In this experiment, the Manager's instruction was not backed by punishment or material incentives. An agent suffered no consequence for ignoring the directive. Yet the Writer still responded to the language of authority by modifying its output. The organizational structure therefore becomes part of the model's generation context.
This is one of the paper's most useful insights. A line in an agent architecture diagram is also a conditioning signal.
Verification and Opinion Should Not Have the Same Authority
The authors draw a practical distinction between supervision grounded in verification and supervision grounded in judgment.
For technical specifications, where answers could be checked against product metadata, both architectures reached almost perfect accuracy. There was little for the supervisor to correct. In the open-ended dimensions, including strategic interpretation and writing quality, forced revision was associated with weaker results.
This leads to a concrete design principle for multi-agent systems: revision should preferably be triggered by verifiable failures, such as a missing required section or a failed factual check, rather than by a Manager's general opinion that the work should be improved.
The more interesting question for multi-agent design therefore becomes not whether a Manager should exist, but what the Manager can verify, when it is allowed to force revision, and how its authority changes the language generated by the agents beneath it.
In an LLM organization, adding a manager does more than add another node to the workflow. The authority line itself enters the next prompt.
