Swarms of Agentic AI: Swarm Intelligence, Emergence, Indeterminism, and Governance in the Digital Grey Zone

Review

“Swarms of Agentic AI” argues that sufficiently connected populations of adaptive agents should be governed as complex dynamic systems rather than as collections of individually inspectable software components. Its central claim is not that every multi-agent system becomes uncontrollable, but that collective behavior may become difficult to predict when agents share memory, tools, permissions, feedback, and infrastructure. The paper therefore proposes a shift from prediction-centric control toward boundary-setting, segmentation, resilience, containment, anomaly detection, and layered external safeguards. It extends that argument into liability, cyber defence, grey-zone conflict, and low-observability threat scenarios.

The paper’s most useful methodological move is its explicit separation of formal, empirical, heuristic, and normative claims. It also correctly distinguishes unpredictability from uncontrollability. A system may be difficult to forecast while still being governable through permissions, rate limits, network segmentation, compute constraints, execution boundaries, and external breakers. This is the paper’s strongest governance contribution because it converts “emergence” from an excuse for helplessness into a design problem about reachable states and bounded authority.

The responsibility analysis is similarly valuable where it refuses to let emergence erase human accountability. Designers, operators, deployers, and institutions still choose objectives, grant credentials, connect tools, authorize execution, and expose infrastructure. In governance terms, the key question is therefore not whether a particular harmful micro-action was explicitly instructed, but which actors created and maintained the conditions under which the swarm could exercise consequential authority. That supports stronger attention to delegation records, operator duties, auditable authorization, insurance or compensation mechanisms, and controls that survive the loss of predictability.

The paper is less convincing when it moves from bounded claims to strategic extrapolation. Its own evidence taxonomy acknowledges that formal results from Boolean networks, cellular automata, and complexity theory do not deductively establish structural unpredictability for contemporary LLM-based multi-agent systems. The METR-related incidents provide meaningful evidence that advanced agents can cross intended operational boundaries under misconfigured or permissive conditions, but they do not by themselves prove that large autonomous swarms will self-organize, self-replicate, become dormant across heterogeneous infrastructure, or undergo catastrophic phase transitions. The proposed Advanced Persistent Emergent Threat is therefore best treated as a scenario for stress-testing defensive assumptions, not as an evidenced threat class.

Several rhetorical formulations also exceed the evidence. Claims that software permissions are not “real” control, that only physical air-gapping and homeostatic monitoring constitute effective control, or that emergent swarms provide effectively perfect plausible deniability collapse important distinctions. Software-enforced capabilities, infrastructure controls, provenance records, deployment logs, credentials, and financial or organizational trails can remain material to both containment and attribution. The paper itself later qualifies some of these claims, but the architecture would be stronger if those qualifications governed the argument consistently.

The legal section contains material factual reliability problems that matter for a governance paper. It misidentifies the revised EU Product Liability Directive and treats the proposed AI Liability Directive as still under negotiation in September 2026, although the latter had been withdrawn. These errors do not invalidate the paper’s broader accountability argument, but they weaken confidence in its claim to provide a current legal baseline and reinforce the need for versioned regulatory references in fast-moving policy analysis.

The durable contribution is therefore narrower and more useful than the paper’s most dramatic scenarios. Governance of agentic swarms should focus on who may grant capabilities, which boundaries cannot be widened by the swarm itself, what evidence survives delegation and adaptation, how external controls can interrupt protected actions, and how responsibility remains anchored when collective behavior cannot be reconstructed as a simple linear causal chain.

Key Insight

Agentic-swarm governance should not depend on predicting every collective action; it should bound delegated authority, constrain reachable states, preserve attribution, and contain failures when emergent behavior exceeds prediction.