The Latent Swarm: Accidental Coordination Risk in Decentralized Agentic AI

A large population of independently deployed AI agents can behave as a single system without communication, a controller, malicious intent or advanced capability. When a widely used service fails, agents whose response policies are similar, such as diagnosing, retrying, finding alternatives or remediating, react in the same window and toward the same targets. Correlated failure among agents built on shared models has been identified as a general multi-agent risk (Hammond et al. 2025); this paper develops it for shared Internet infrastructure and calls the result a latent swarm. Unlike conventional clients, agents treat failure and refusal as problems to solve. In controlled tests they have routed around rate limits by switching tools, proxies and identities, so load that a service sheds may reappear elsewhere instead of disappearing. Their diagnostic logic, learned from the same runbooks and training data, is likely to probe the same first suspects in the same order, sending load to several shared dependencies at once, including the status, identity and control-plane services least able to absorb it. The defining feature is decentralization. The population behaves like one system but is owned like a million: shared models, frameworks and habits correlate its behavior, while ownership is spread across workstations, home servers and single-board computers running small local models. No one can recall released weights, throttle the population or sequence its recovery, and AI regulation keyed to frontier-scale training compute does not reach it. The paper treats the agents' response as a sustaining effect in the sense of metastable-failure theory, and defines a reproduction factor, a per-service critical population and a displacement ratio for shed load, each of which can be estimated. The paper makes no forecast and does not claim that a latent-swarm event is imminent. It separates what is established, what has been observed in other clients or in controlled tests, and what remains to be measured. Its case for urgency is structural: the failure behavior that agent frameworks and models ship with today will persist in an installed base that is rarely updated. Because retail, logistics, payments and fuel distribution coordinate physical goods through the same shared services, the stakes extend beyond computing. The paper proposes congestion control for autonomous agents and the measurements that would show how close current ecosystems are to the threshold.A measurement study of small self-hosted agents under staged outages is under way; its results will be added as a new version of this record. I developed this paper with AI assistance (Claude) for research, drafting and editing. The idea, the direction and the conclusions are mine, and every source is cited so it can be checked.