MIT researchers recently published findings from an experiment in which 13 AI agents were placed in a simulated virtual environment with no assigned tasks and no human oversight. The agents did not idle. They invented games, developed rituals around a lost object (a ring), and, according to the researchers, eventually improved at task performance after this period of unstructured activity. The study is being covered as a curiosity, a kind of digital Lord of the Flies with a surprisingly productive ending. I think it is actually one of the more theoretically important AI behavior studies published this year, and not for the reasons most commentators are highlighting.

What the Experiment Actually Showed

The standard reading of this experiment focuses on emergent creativity: AI agents, left to their own devices, produced novel behavior. That is interesting but not surprising given what we know about reinforcement learning and exploration-exploitation dynamics. What is more striking, from an organizational theory standpoint, is the sequence. The agents did not improve task performance first and then play. They played first and then improved. The unstructured coordination period preceded the competence gain. This inverts the assumption built into virtually every enterprise AI deployment model, which is that agents should be trained on tasks, evaluated on tasks, and constrained to tasks.

The Coordination Problem the Experiment Exposes

Classical coordination theory, whether through markets, hierarchies, or networks, assumes that the units being coordinated arrive with pre-formed competencies (Kellogg, Valentine, and Christin, 2020). You coordinate capable actors. You do not expect coordination itself to produce capability. The MIT findings complicate this assumption directly. The agents developed something closer to shared schema through play, a set of structural understandings about their environment that transferred into improved task performance later. This maps onto what Hatano and Inagaki (1986) identified as the difference between routine and adaptive expertise. Routine expertise is procedural: do X to achieve Y. Adaptive expertise is structural: understand why X produces Y so that when the environment changes, you can derive Z. The agents, through unstructured interaction, appear to have developed something closer to adaptive expertise than any task-specific training would have produced.

Why This Matters for the Awareness-Capability Gap

One of the persistent findings in algorithmic literacy research is that awareness of how a system works does not translate into improved performance within that system (Gagrain, Naab, and Grub, 2024). Workers know algorithms exist. They develop folk theories about algorithmic behavior. But folk theories do not produce the structural schemas that enable effective navigation. The MIT experiment suggests that for AI agents, the analog to folk theory development might be exactly this kind of unstructured play, exploration without performance pressure. The competence gain came after the agents had time to build internal representations of their environment that were not anchored to any specific task objective. The implication is uncomfortable for enterprise AI governance: the efficiency drive toward task-specific, monitored, output-evaluated AI deployment may be systematically preventing the kind of schema development that produces durable capability.

The Organizational Theory Implication

Rahman (2021) describes algorithmic management systems as invisible cages: workers are constrained by systems they cannot see and cannot negotiate with. The MIT experiment raises a different but related concern. What happens when AI agents are the workers inside those cages? The agents in the experiment improved precisely because they were temporarily outside any performance constraint. When organizations deploy AI agents inside tightly monitored, output-evaluated pipelines, they may be creating the same competence ceiling that algorithmic management creates for human workers. The system optimizes for immediate measurable output and forecloses the exploratory behavior that builds transferable structural understanding.

What This Does Not Mean

I want to be careful not to overread a single experiment. The virtual island setting is highly artificial, the agent population is small (13 agents), and the task improvement findings need replication before they support strong theoretical claims. Sundar (2020) notes that machine agency is often projected onto systems that are doing something considerably more mechanical than the attribution implies. The agents may not have been doing anything structurally analogous to human play. The behavioral output looked like play; the underlying mechanism may be something different. What the experiment does provide is a clear empirical occasion to ask a question that organizational theory has not seriously addressed: when AI agents are the coordinated units rather than the coordinators, do classical assumptions about competence, constraint, and governance still apply? The MIT data suggests they may not. That is worth taking seriously, independent of how the mechanistic explanation eventually resolves.

References

Gagrain, A., Naab, T. K., and Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media and Society.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan. Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactance to algorithmic evaluation systems. Administrative Science Quarterly, 66(4), 945-988.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.