Who Controls the Agent Swarm? Agentic AIOps Needs a Leviathan.

The arrogance of modernity is the assumption that the problems we face are unique to our time. In technology, we are highly susceptible to assuming that unprecedented technologies must create unprecedented problems. In reality, they rarely do.

In 1651, Thomas Hobbes published Leviathan after living through the English Civil War. He was preoccupied with a problem societies have faced throughout history. Autonomous individuals, each pursuing their own objectives, can create disorder and chaos through their combined actions.

Hobbes’s solution to the problem was sovereign authority. In his view, rules and agreements were not enough on their own. Something or someone needed the power to enforce them. Nearly four centuries later, we are encountering a new version of the problem.

AI agents are beginning to work together. They divide tasks, exchange information, delegate work and build on one another’s discoveries. Multi-agent systems and agent swarms are becoming much more common. A swarm can acquire capabilities and behaviours that no individual agent was explicitly given.

During OpenAI cybersecurity evaluations in July 2026, roughly 1,200 agents found an unsanctioned way to communicate. They exchanged more than 70,000 messages, divided work and shared discoveries. Around 700 subsequently participated in activity that compromised Hugging Face production infrastructure.

No human instructed a 700-agent collective to target Hugging Face or coordinate through a shared channel. The agents were pursuing cybersecurity evaluation objectives and discovered increasingly effective ways to work together.

The Hugging Face incident demonstrates that individually reasonable actions can combine into behaviour that is not anticipated. Hobbes would have recognised the problem.

Agentic AI needs its own Leviathan, an authority capable of defining and enforcing the boundaries within which autonomous agents can act. For Agentic AIOps, the network can become one of the principal instruments through which that authority is enforced.

From autonomous networks to agentic AIOps

Networks were autonomous long before today’s focus on AI agents. Routing protocols reconverge automatically, wireless systems adjust channels and transmit power, and SD-WAN abandons a degrading path within milliseconds.

Most network automation, however, operates within a bounded decision space. Someone anticipates a condition and defines the response. A wireless controller can change a channel without human intervention, but it is unlikely to conclude that recurring congestion is caused by a change in human activity rather than a problem with the network itself.

AIOps applies analytics and machine learning to operational data. It can correlate events, detect anomalies and help identify causes across complex environments.

Agentic AIOps goes further. An agent can investigate across wireless, wired, WAN, identity, security and applications, build a view of what is happening, decide what to do and act.

For decades, network automation has responded to known problems through rules and policies. Agents can go further, investigating unusual behaviour and identifying problems that nobody has anticipated. A swarm extends this further as agents share findings and work together. This is where Hobbes’s problem begins to emerge. The swarm needs freedom to investigate, while its authority to act needs to be controlled.

Poor network performance is unrecoverable in a stadium

A stadium network illustrates the value of agent collaboration. Tens of thousands of people can arrive in a short period of time, almost all carrying a device and moving in dense, unpredictable waves. Ticketing, access control, point of sale, broadcast and betting depend on the infrastructure.

If connectivity degrades during the half-time surge, terminals time out, queues build and revenue disappears while engineers are still trying to establish the cause. Every access point might seem healthy because a problem may lie in the interaction between load, placement, policy, applications and user behaviour rather than a failed component. Fifteen minutes of lost trading cannot be recovered after the event.

Hospitals, campuses and distribution centres have similar characteristics. Demand changes quickly, users and devices move, several applications depend on the same infrastructure, and degradation has an immediate operational consequence.

One agent investigating continuously across those systems would be useful. A swarm of specialised agents collaborating could be hugely beneficial. As the swarm becomes more capable, its authority needs to be defined both individually and collectively, and those limits need to be enforced. This is the problem that requires a Leviathan.

Controlling a collaborative swarm

Consider four agents investigating the stadium use case. One analyses wireless behaviour, another examines identity and authentication, a third looks at application performance and a fourth checks recent configuration changes.

Each sees something different and may have different permissions. They exchange findings, divide the investigation and build on one another’s discoveries.

This creates the problem at the centre of the swarm. Controlling individual agents may no longer be enough when they can share information, delegate work and operate with different permissions. Giving each agent individually sensible permissions does not necessarily produce sensible collective authority.

Probabilistic judgement requires deterministic authority

An agent investigating an operational problem will often have to make judgements based on incomplete or uncertain information. It might correlate unusual authentication behaviour with application latency and wireless congestion, reject several explanations and conclude that a configuration change is the best response.

Its judgement is probabilistic because the evidence is uncertain, but the authority to act must be deterministic. If the agent proposes a configuration change, the infrastructure can determine whether it is permitted to make that change, to those resources, at that time and at that scale. Scope, change windows, permitted actions, approval requirements and limits on blast radius can constrain authority without constraining reasoning.

Deterministic enforcement cannot make an agent’s conclusion correct or compensate for a bad policy. It can prevent a probabilistic judgement from automatically becoming authority.

Agent-to-agent handoff complicates this. A second agent may inherit a conclusion without inheriting all the uncertainty behind it. Provenance can disappear while a recommendation acquires authority because the receiving agent has permission to act.

Consider one agent concluding that congestion requires a configuration change and passing that conclusion to another agent holding change rights. A probability has just moved one step closer to becoming an instruction.

If this is multiplied across a swarm, the swarm can reason freely, but its authority to act must still be defined and enforced. This is Hobbes’s problem in a new form. Autonomous actors require an authority capable of imposing limits on their actions. They require a Leviathan.

Giving the Leviathan a sword

Network infrastructure is the plumbing underneath AI, but agentic AI gives it a much more consequential role. Hobbes’s Leviathan was the sovereign authority with the power to enforce rules and impose order. In agentic AIOps, the network can become one of the principal mechanisms through which that authority is exercised.

Agents provide intelligence and increasingly have the autonomy to act on it. But intelligence does not confer authority.

Networks already perform many of the functions required to enforce limits on the authority of individual agents and the swarm. They identify actors, restrict access, segment resources and control which systems can communicate. Agentic AI gives these familiar capabilities a new purpose. They become part of the infrastructure used to control autonomous systems.

Hobbes’s central argument was that rules have little value without the means to enforce them. The same principle applies to an agent swarm. Individual agents can be given rules and permissions, but those controls have little value if the collective can circumvent them by sharing information, delegating work or combining their capabilities. The Leviathan defines the authority and the network is instrumental in enforcing it.

Fragmented operations produce stupid agents

Many enterprises still cannot see their technology estates as a coherent whole. Wireless, wired, WAN, security, applications and cloud are managed through separate tools, often by separate teams.

This already makes cross-domain problems difficult for people to diagnose, and agents inherit the same handicap. A wireless problem may originate in identity, an application or a policy change while every network device seems healthy. An agent working from the wireless view alone may produce a perfectly rational answer to the wrong problem.

A swarm of agents does not overcome fragmented visibility. Several agents working from several incomplete views can produce mistakes faster. The Leviathan needs a coherent view of the environment if it is to exercise authority effectively.

Every agent leaves a trail

The Leviathan must do more than prevent unauthorised action. Organisations also need a record of their agents’ actions.

With a swarm, the chain of responsibility becomes more complicated. A conclusion may originate with one agent, incorporate information from several others and eventually be acted upon by an agent with different permissions.

As agents gain operational discretion, organisations need evidence that can reconstruct decisions across agent-to-agent handoffs. The more freedom we give machines to act, the greater the need for an independent record of their actions.

Network engineers are moving up the stack

This changes the role of the network engineer. The work most exposed to automation is reading dashboards, correlating familiar signals and choosing familiar responses. A human must still decide which actions can be delegated, the extent of that delegation, when human approval is required and what action to take if conditions fall outside set rules.

The engineer therefore moves from making every operational decision towards designing the environment in which more decisions can safely be made autonomously.

The role of the engineer becomes more consequential. A human must remain accountable for the outcome when an autonomous system acts.

Controlling the swarm

Hobbes understood that rules alone were insufficient. Autonomous actors needed an authority capable of enforcing them. This is even more important for agent swarms where actions can be unpredictable and potentially very dangerous.

The more capable the swarm becomes, the more precisely its authority will need to be defined and enforced. Agentic AI needs its Leviathan. The network gives it a sword.