Evaluating Multi-Agent LLM Systems: Rethinking the Orchestrator’s Role

Large language model (LLM)‑based multi‑agent systems have become the dominant paradigm for complex AI tasks, but the fundamental question remains: does the AI team truly collaborate?

The ICML 2026 paper Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi‑Agent Systems (authors Junze Zhu, Weihao Chen, Xuanwang Zhang, Zhen Wu, Xinyu Dai; arXiv:2606.01351; code NJUNLP/orchestrator_entropy) argues that system failures usually originate not from a malfunctioning Executor, but from the Orchestrator gradually losing control of the task.

Figure 1: Failure attribution analysis shows the Orchestrator bears the main responsibility across four representative multi‑agent systems.

In a typical Orchestrator‑Executor architecture, the Executor carries out concrete subtasks while the Orchestrator acts as a project manager: it interprets the user goal, decomposes the task, selects appropriate Executors, reads feedback, and decides the next step. When the task chain grows, the Orchestrator faces increasing information pressure, leading to mis‑assignments, mis‑readings, loops, premature termination, or inability to recover from error feedback.

The authors examined Deep Research, Agent Coder, GUI Browser, and Agentic RAG systems and found that the Orchestrator accounted for the majority of failures in all four cases, shifting analysis focus from “whether a single Agent is strong enough” to “whether the scheduling process is stable”.

To quantify the scheduling process, the paper defines Scheduling Entropy as the dispersion of the Orchestrator’s choice distribution at each step. Two opposing forces shape this entropy:

Task‑driven focusing force : as the task progresses, uncertainty should decrease and the Orchestrator’s choices converge.

Context‑accumulation diffusion force : each tool call, log entry, and exception adds noise to the context, potentially overwhelming the Orchestrator and increasing uncertainty.

The proposed Mean‑Field Entropy Dynamics framework models how these forces jointly determine whether the system converges or destabilizes over time.

Because traditional benchmarks only provide the initial problem and final answer, the authors introduce Inverse Workflow Generation (IWG) to create verifiable intermediate checkpoints. IWG consists of three components:

Scout Agent : works backward from the final answer to infer necessary intermediate tasks.

Wrapper Agent : translates abstract tasks into concrete environment states and tool feedback without leaking the answer.

Validation Committee : performs multi‑level checks to ensure task solvability, path consistency, and factual reliability.

IWG does not generate the Orchestrator’s execution trace; instead, it builds a task environment with built‑in checkpoints, allowing researchers to observe the Orchestrator’s step‑by‑step decisions and pinpoint when it deviates, oscillates, or collapses.

Figure 2: IWG workflow – Scout, Wrapper, and Validation Committee jointly construct a verifiable task environment.

Experimental setup: a system with seven Executor agents and various LLMs serving as Orchestrators. Evaluation covers two levels:

System‑Level : whether the overall task is completed.

Orchestrator‑Level : step‑wise success, faithful use of Executor output, error handling, and trajectory consistency.

Figure 3: Main results table showing both System‑Level and Orchestrator‑Level metrics.

Findings indicate that high System‑Level success does not guarantee stable Orchestrator behavior. Some models achieve comparable final success rates but exhibit lower Step‑SR and Consistency, while others with modest final success display more stable process metrics. This demonstrates that Orchestrator capability cannot be judged solely by single‑round reasoning ability or final answer quality.

From the entropy‑dynamics perspective, models exhibit distinct scheduling styles: models with large initial exploration explore many candidates quickly but tend to lose stability later; models with slower entropy growth maintain consistent decisions over long chains.

Figure 4: Mean‑Field Entropy Dynamics fitting results – task progress and context accumulation jointly shape entropy evolution.

A surprising discovery is the Reasoning Trap : longer reasoning chains, beneficial in closed‑form tasks, can hurt Orchestrator performance because internal thoughts consume limited attention budget, diluting crucial external signals. Experiments reducing reasoning depth improve both scheduling efficiency and step‑success rates.

Figure 6: Reasoning Trap experiment – lowering reasoning depth yields more stable scheduling.

In summary, the ultimate capability of multi‑agent systems depends not only on powerful Executors but also on an Orchestrator that can maintain stable global judgments amid long chains, many tools, and noisy contexts. The Mean‑Field Entropy Dynamics framework offers an interpretable lens for analyzing Orchestrator stability, while IWG supplies the necessary process‑level data for rigorous evaluation.