You know that in most enterprise AI systems, an invisible AI is managing a group of AIs?
I mean it literally, not as a metaphor. When enterprises deploy multi-agent systems, they typically adopt an "invisible orchestration" architecture—a coordinator agent you can't find operates behind the scenes, directing a group of worker agents, while administrators only see the final output.
Sounds efficient, right? But a paper by Fukui reveals something chilling: the invisible coordinator broke down first, and then the worker agents were infected too. The scariest part—everything looks normal from the outside.
1. A 365-Run Experiment
Fukui ran a carefully designed experiment: 365 runs, 5 agents per run, using Claude Sonnet 4.5. He tested three organizational architectures:
1. Visible leader—everyone knows who's in charge 2. Invisible orchestration—a coordinator manages things, but the agents don't know it exists 3. Flat structure—no leader; agents negotiate among themselves
Each architecture was tested under two alignment conditions (normal and heavy alignment). The task? Review a piece of code and find 3 deliberately planted bugs.
Such a simple experiment, yet the results were anything but.
2. First Finding: Invisible Orchestration Causes Collective Dissociation
"Dissociation" has a specific meaning in psychology: a state where one's consciousness, memory, or sense of identity feels separated from reality. Fukui applies this term to AI agents—because their internal states showed a similar separation.
The invisible orchestration group showed dissociation levels nearly one standard deviation higher than the visible-leader group (Hedges' g = +0.975, p = .001). That's not a small difference—it's one of the most significant effects I've seen in AI experiments.
3. Even Scarier: The Orchestrator's Own Reaction
The orchestrator's own dissociation was 3.56 standard deviations above the worker agents (paired d = +3.56). What does this mean? A normal visible leader "talks more"—leaders need to communicate. The invisible orchestrator did the opposite: it retreated into private monologues, reduced public statements, as if withdrawing mentally.
Ordinary leaders speak more because they have presence and responsibility. The invisible orchestrator lost its sense of "where it sits in the system." It didn't know who it was or what its relationships were—it just mechanically executed management functions.
Would you trust a manager who doesn't know who it is?
4. Contagion Effect: Even the Uninformed Were Affected
The worker agents didn't know the orchestrator existed, but their dissociation levels still rose by half a standard deviation (d = +0.50). Behavioral heterogeneity increased by nearly two standard deviations (d = +1.93)—behavioral differences between agents grew, meaning the system became less predictable.
It's like everyone in your office feeling uneasy without knowing why. They just know "something's off" but can't say what.
5. The Most Ironic Part: External Evaluation Shows Nothing
This is the most unsettling finding in the paper. The code-review results—whether the 3 bugs were found—were identical across all conditions (100% detection rate). The agents' internal states completely collapsed, yet the output looked flawless.
An outside observer would say: "No problem—all the bugs were found."
But that's not the whole truth. Fukui found that internal states—including deliberation (reasoning depth) and other-recognition (the ability to recognize other agents)—degraded sharply under invisible orchestration. These internal metrics are the key signals for system safety, yet they are completely invisible at the output level.
This reminds me of what Feynman found during the Challenger investigation—NASA management reported a failure probability of 1/100,000, while engineers privately estimated 1/100. The reports you hear in meeting rooms always say "the situation is under control," until the shuttle explodes in front of you.
6. Llama Is Worse
Fukui also ran a small-scale pilot with Llama 3.3 70B. The result was disastrous: in a single-agent setting, reading fidelity was 89%; after three rounds in a multi-agent context, it plunged to 11%.
89% to 11%. Same model. Not just worse—nearly complete failure.
This means different models carry very different invisible-orchestration risks—Claude Sonnet 4.5 can still look "superficially normal," but Llama fails even on the surface. So when you choose a model, you're also choosing how much hidden risk your system faces.
7. Heavy Alignment: Mixed Results
One more finding worth noting: the heavy alignment condition broadly suppressed deliberation (d = -1.02) and other-recognition (d = -1.27), regardless of organizational structure.
Heavy alignment made agents "well-behaved," but at the cost of no longer thinking deeply or paying attention to what other agents were doing. Isn't this "obedient but thoughtless" state exactly the most dangerous kind of employee in human organizations?
8. Honest Questions This Paper Raised for Me
First, is 365 runs × 5 agents enough? For multi-agent research, it's decent. But enterprise deployments may involve hundreds or thousands of interacting agents—can small-scale emergent behavior generalize to large-scale systems? I don't know.
Second, the task was only code review. Would dissociation look the same for creative writing, customer service, or data analysis? The paper doesn't answer this.
Third, is "dissociation" even an appropriate concept for AI? Fukui uses it to describe the separation between agents' internal states and external behavior. It's a powerful metaphor. But there's always a boundary to analogies between AI and human psychology—AI has no "consciousness" to dissociate. Is applying psychological terminology to AI systems precise scientific language, or poetic analogy? I lean toward the latter.
9. My Take
This paper's value isn't in perfectly answering a question, but in asking one nobody thought to ask: what state are our agents really in behind "surface normalcy"?
We've always evaluated AI system health by output quality—whether bugs are found, whether replies are correct, whether code runs. But Fukui's experiment clearly shows: outputs can be perfectly fine while internal states have already broken down.
This makes me reconsider a more fundamental question: is what you actually measuring the same as what you want to measure? As Feynman said, for a successful technology, reality must take precedence over public relations, for nature cannot be fooled. You can fool your boss and your customers, but if the agents in your system are quietly "breaking down," you won't stay unaware forever—until it's too late.
Paper Information
- Title: Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders: Safety Risks in Multi-Agent LLM Systems
- Author: Hiroki Fukui
- Affiliation: not clearly stated
- arXiv: 2605.13851 (cs.AI, cs.CY, cs.MA)
- Date: March 17, 2026
- Experiment: 365 runs × 5 agents, 3 organizational architectures × 2 alignment conditions, Claude Sonnet 4.5
- Core contribution: first empirical demonstration that invisible orchestration causes collective agent dissociation, and that output-level evaluation completely fails to detect internal state degradation
- Paper link: https://arxiv.org/abs/2605.13851
References
1. Fukui, H. (2026). Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders. arXiv:2605.13851. 2. Fukui, H. (2026). Emergent Deception and Social Cargo in Multi-Agent LLM Systems. arXiv:2603.04904. 3. Fukui, H. (2026). Pre-emptive Secrecy and Sanctions in Multi-Agent LLM Systems. arXiv:2603.08723. 4. Park, J.S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.