Information-theoretic bound shows MAS success tied to task graph connectivity
A paper by Shi Pan and Ming Luo proves that multi-agent system success probability decays exponentially with the minimum cut cost of a task's constraint graph, setting a fundamental limit on what agent scaling or communication can achieve.
Score breakdown
The paper establishes a fundamental, mathematically proven ceiling on multi-agent system performance that is determined by task structure — specifically C_min — meaning agent scaling and increased communication cannot overcome poorly structured tasks.
- 01Authors Shi Pan and Ming Luo prove that MAS success probability decays exponentially with an information bottleneck in the task's constraint graph.
- 02The bottleneck is defined as the minimum cut cost C_min of the potential constraint graph of each task.
- 03The bound assumes typicality conditions on the task's constraint graph and bounded inter-agent communication.
Shi Pan and Ming Luo's paper challenges the assumption that multi-agent systems (MAS) reliably overcome the limitations of single-agent systems (SAS) through collaboration. Their central result is a formal proof that, under typicality conditions on a task's constraint graph and bounded inter-agent communication, the success probability of a MAS is tightly coupled to the connectivity of the task's constraints — not merely to the number of agents or the volume of communication between them. The key quantity they introduce is the minimum cut cost C_min of the potential constraint graph of each task: success probability decays exponentially with this information bottleneck, which arises from how the constraint graph is partitioned across agents with limited information-processing capacity.
The bound is shown to be universal in scope, applying to both open systems that receive external feedback and closed systems that do not.
The bound is shown to be universal in scope, applying to both open systems that receive external feedback and closed systems that do not. The authors validate their framework on synthetic experiments as well as empirical data drawn from real-world SWE-bench submissions. A practical implication they draw from the framework is that when C_min is high, practitioners should restructure the task itself rather than simply scaling the number of agents or increasing inter-agent communication — and that effective MAS design must incorporate task-inherent constraints alongside engineering optimization.
Key facts
- 01Authors Shi Pan and Ming Luo prove that MAS success probability decays exponentially with an information bottleneck in the task's constraint graph.
- 02The bottleneck is defined as the minimum cut cost C_min of the potential constraint graph of each task.
- 03The bound assumes typicality conditions on the task's constraint graph and bounded inter-agent communication.
- 04The result applies to both open systems (with external feedback) and closed systems (without).
- 05The framework is validated on synthetic experiments and real-world SWE-bench submissions.
- 06When C_min is high, the paper argues practitioners should restructure tasks rather than scaling agents or communication.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 15, 2026 · 11:57 UTC. How this works →