Executive Summary
This paper presents empirical evidence that organizational dysfunction is substrate-independent. In a controlled comparison, four coordination architectures (single agent, hierarchical, stigmergic with 8 concurrent agents, and gated pipeline) built the same 7-service backend using the same LLM and a $50 budget. Performance was inversely correlated with coordination complexity: 28/28, 18/28, 9/28, and 0/28. The pipeline spent its entire budget on planning, the hierarchical coordinator refused to delegate, and the stigmergic agents produced incompatible interfaces at every boundary. Only the single agent succeeded fully.
In two further studies, a pipeline swarm equipped with six explicit anti-dysfunction mechanisms produced the dysfunction those mechanisms were designed to prevent: bikeshedding, governance conflicts, backward pipeline oscillation, and verification theater. A contract-first alternative that replaces subjective evaluation with mechanical test verification narrowed the Goodhart gap but introduced its own dysfunction (specification perfectionism), suggesting that dysfunction migrates across architectures rather than disappearing.
Key Contributions and Methodology
First, the controlled architecture comparison and the two follow-up studies, which document recognizably organizational failure modes in LLM agent teams.
Second, a token-level decomposition across five communication topologies (Centralized, Fully Connected, Hierarchical, Ring, Star-Mesh). Governance overhead scales with communication links: Fully Connected governance grows from 25.5% at n=3 to 69.4% at n=15, while Centralized grows by only 0.51% per agent, reflecting O(n²) versus O(n) link complexity. The model's prediction for a 5-agent Centralized pipeline (24.0%) is consistent with the 22.6% governance overhead measured in Study 1.
Third, an information-theoretic treatment of each coordination stage as a lossy channel. Intent entropy degrades monotonically through the pipeline, as the Data Processing Inequality predicts: the single agent retains 85.8% of the original specification entropy at verification, the hierarchical architecture only 25.8%. Crawford-Sobel strategic communication explains the per-layer loss.
Key Findings
- Architecture comparison: 28/28 (single agent), 18/28 (hierarchical), 9/28 (stigmergic), 0/28 (gated pipeline) on the same task, model, and budget
- Overhead scaling: Fully Connected governance rises from 25.5% (n=3) to 69.4% (n=15); Centralized grows at 0.51% per agent
- Structural invariance: governance tokens have a coefficient of variation of 0.0000 across capability levels; as agents improve, the governance fraction increases because implementation cost shrinks while coordination cost stays constant
- Entropy loss: 85.8% of specification entropy retained by a single agent versus 25.8% by the hierarchical architecture
- Dysfunction migrates: anti-dysfunction mechanisms and contract-first verification shift the failure mode rather than eliminating it
Key References
Strategic Information Transmission. Econometrica, 50(6), 1431-1451.
Elements of Information Theory (2nd ed.). Wiley-Interscience.
Why Do Multi-Agent LLM Systems Fail? NeurIPS 2025 Datasets and Benchmarks Track. arXiv:2503.13657.
A Theory of "Yes Men". American Economic Review, 83(4), 757-770.
Problems of Monetary Management: The U.K. Experience. Papers in Monetary Economics, 1.