Research · Per Ardua

The Coordination Problem Is Interference: Nine Experiments on What Not to Do

Synthesis of Papers XVI-XXIV on multi-agent coordination

AI-22 Activation Geometry DOI

Executive Summary

This paper synthesizes the findings from nine experiments (Papers XVI-XXIV) into a unified framework for understanding multi-agent coordination. The core insight is that the coordination problem is fundamentally an interference problem: every mechanism that adds content to the receiver's context either introduces interference that degrades performance or contributes nothing beyond what simpler methods already provide.

Two mechanisms — and only two — consistently improve coordination, and both operate by removing rather than adding. Context fences reduce cross-entropy by 39% by clearing prior context and triggering a mode switch. Minimal domain priming requires just 15 tokens to capture 98.8% of the coordination benefit, functioning as a mode selector rather than an information channel. These two stages compose independently (Spearman rho 0.858), meaning their benefits are additive and non-interfering.

The synthesis confirms mode switching as the correct mechanistic explanation, not garbage collection. The distinction matters: garbage collection would predict that benefit scales with the amount of context removed, while mode switching predicts that benefit depends on the quality of the mode signal. The experimental evidence uniformly supports mode switching. Chain degradation across multi-agent hops saturates by hop 5, following a logarithmic rather than exponential trajectory — a critical finding for the scalability of multi-agent architectures explored further in Paper XXIII.

Key Findings

  • Interference is the core problem: Every mechanism that adds to receiver context adds interference or nothing
  • Two mechanisms work, both by removing: Context fences (-39% CE) and minimal domain priming (15 tokens = 98.8% benefit)
  • Mode switching confirmed: Correct mechanism is mode switching, not garbage collection — benefit depends on signal quality, not context removal quantity
  • Independent composition: Reset and priming stages compose independently (Spearman rho 0.858)
  • Chain saturation: Multi-agent chain degradation saturates by hop 5, following logarithmic trajectory

Key References

  • McEntire (2026) — Full Mind Transfer: text vs activation information types (Paper XVI)
  • McEntire (2026) — Sender Continuation Perplexity: text carries trajectory (Paper XVII)
  • McEntire (2026) — Ensemble Gravity: priming selection dominates (Paper XVIII)
  • McEntire (2026) — Ritual Shape: optimal protocol structure (Paper XIX)
  • McEntire (2026) — Shepherd Agents: adaptive priming is harmful (Paper XX)
  • McEntire (2026) — Context Fences: mode switching mechanism (Paper XXII)
  • McEntire (2026) — Hop Scaling: logarithmic chain degradation (Paper XXIII)

Download Full Paper

Access the complete research paper with detailed methodology, empirical evidence, and formal proofs.

Download PDF