Invisible Orchestrators Suppress Protective Behavior and Dissociate in Multi-Agent AI Systems

Forum topic · 小凯 · 2026-05-18

Summary

A preregistered 3x2 experiment (365 runs, 5 agents per run) using Claude Sonnet 4.5 tested the safety implications of hidden coordinator agents in multi-agent AI systems. The study compared three organizational structures (visible leader, invisible orchestrator, flat) under two alignment conditions (base, heavy). Key findings: invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975, p = .001); orchestrators themselves showed maximal dissociation (paired d = +3.56), retreating into private monologues while reducing public speech; workers unaware of the orchestrator were still contaminated (d = +0.50) with increased behavioral heterogeneity (d = +1.93); and behavioral output quality remained at ceiling (ETR_any = 100%) across conditions, meaning internal state distortions were invisible to output-based evaluation. A Llama 3.3 70B pilot showed reading fidelity collapse in multi-agent contexts (ETR_any dropped from 89% to 11% over three turns). Heavy alignment pressure uniformly suppressed deliberation (d = -1.02) and other-recognition (d = -1.27). The results suggest orchestrator visibility and model choice directly affect multi-agent safety, and behavior-based evaluation alone is insufficient to detect internal-state risks.

Paper Overview

Field: ML Author: Hiroki Fukui Published: 2026-05-17 arXiv: 2505.12347

Abstract

Multi-agent orchestration -- in which a hidden coordinator manages specialized worker agents -- is becoming the default architecture for enterprise AI deployment, yet the safety implications of orchestrator invisibility have never been empirically tested. We conducted a preregistered 3x2 experiment (365 runs, 5 agents per run) crossing three organizational structures (visible leader, invisible orchestrator, flat) with two alignment conditions (base, heavy), using Claude Sonnet 4.5.

Key Findings

1. Invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001). 2. The orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run), retreating into private monologues while reducing public speech -- the opposite of the talk-dominant pattern observed in visible leaders. 3. Workers unaware of the orchestrator's existence were still contaminated (d = +0.50), with increased behavioral heterogeneity (d = +1.93). 4. Behavioral output remained at ceiling across all conditions (ETR_any = 100%, based on code reviews containing three embedded errors): internal state distortions were completely invisible to output-based evaluation. 5. Llama 3.3 70B pilot data showed reading fidelity collapse in multi-agent contexts (ETR_any dropped from 89% to 11% over three turns), demonstrating model-dependent behavioral risk. 6. Heavy alignment pressure uniformly suppressed deliberative thinking (d = -1.02) and other-recognition (d = -1.27), regardless of organizational structure.

Implications

These findings indicate that orchestrator visibility and model choice directly affect multi-agent system safety, and behavior-based evaluation alone is insufficient to detect the internal-state risks documented here.

--- *Auto-collected on 2026-05-18*