The Organizational Physics of Multi-Agent AI: Substrate-Independent Dysfunction in Autonomous Software Engineering Swarms

Executive Summary

The emerging paradigm of multi-agent AI systems — autonomous software engineering swarms, collaborative reasoning frameworks, and distributed problem-solving architectures — is celebrated as a path to scalable intelligence. This paper demonstrates that these systems are subject to the same organizational physics that constrain human organizations. Coordination overhead scales with the number of communication links, not with the capability of individual agents. Increasing model intelligence does not reduce coordination costs; it merely changes the substrate on which those costs are incurred.

Token governance analysis across multi-agent configurations reveals that overhead consumes 25.5% of total token budget at n=3 agents, rising to 69.4% at n=15 agents. Critically, the coefficient of variation across model scales is 0.0000, meaning that overhead percentages are identical whether the agents are small or frontier-scale models. This structural invariance directly addresses the "frontier models fix this" argument: they do not, because the overhead is a property of the communication topology, not of individual agent capability.

The paper develops a formal information-theoretic model connecting multi-agent coordination loss to the Strategic Rate-Distortion-Perception framework. The model predicts 24.0% information loss at n=3, closely matching the 22.6% empirically observed in controlled studies. This calibration validates the formal model and enables prediction of coordination costs for configurations not yet tested, providing a principled basis for deciding when multi-agent architectures provide net benefit versus net overhead.

Key Contributions and Methodology

The paper makes three contributions. First, it provides the first systematic measurement of coordination overhead in multi-agent AI systems as a function of both team size and model capability. By controlling for task complexity and measuring overhead in tokens (the native unit of computation for language models), the analysis reveals the scaling law governing multi-agent coordination: overhead is quadratic in team size (scaling with communication links) and invariant to model scale.

Second, it develops a formal information-theoretic model of multi-agent coordination loss. Each agent-to-agent communication channel introduces lossy compression (finite context windows), strategic distortion (divergent sub-objectives), and reconstruction error (misaligned representations). The aggregate information loss is modeled as a function of channel count, channel capacity, and preference divergence, connecting directly to the Strategic RDP framework developed in the companion paper.

Third, the paper addresses the structural invariance argument head-on. The "frontier models fix this" claim implicitly assumes that coordination overhead is a capability limitation that scales away. The data show otherwise: overhead percentages are identical across model scales (CV=0.0000), because the overhead arises from the communication topology — the number and structure of inter-agent channels — not from the intelligence of the agents using those channels. Smarter agents communicating through the same topology incur the same proportional overhead.

Key Findings

• Overhead scaling: Coordination overhead rises from 25.5% (n=3) to 69.4% (n=15), scaling with communication links rather than linearly with team size
• Scale invariance: Overhead percentages are identical across model scales (CV=0.0000), refuting the claim that more capable models reduce coordination costs
• Model calibration: Information loss model predicts 24.0% at n=3, closely matching the 22.6% observed in experimental Study 1
• Formal IT foundation: Coordination loss is connected to the Strategic RDP framework, enabling principled prediction of overhead for untested configurations
• Substrate-independence: The same organizational physics that constrain human teams — communication overhead, information loss, strategic misalignment — apply identically to AI agent swarms

Key References