The Rise of AI Agents Increases the Risk of Losing Control

Companies are deploying fleets of interconnected agents that call APIs and access applications not designed for them. The major challenge lies in the chain of interactions, which is rarely mapped. A unique identity, real-time supervision, and preventive application are necessary to avoid losing control.

Deficient Traceability and Diluted Responsibilities

Enterprise AI programs often stagnate when human managers lose visibility into the actions of their agents. Security teams struggle to determine which agents have access to which systems and to trace which agent triggered an action downstream several steps earlier. A support ticket can pass through four agents before being seen by a human, with each transition adding a decision point that may not have been approved. As the processing chain lengthens, responsibility dilutes: if five agents are involved and a break occurs at step four, it becomes difficult to identify who is responsible, especially since organizational charts often stop at deployment without designating a human respondent. To maintain control, organizations must be able to answer at any moment the question: what is the system doing and who is accountable for it?

Governing the Chain: Identity, Supervision, Application

The first step is to give each agent a unique identity, distinct from the deployer's permissions: a name in a registry, a defined authority, and a clearly designated human sponsor. This foundation is essential but not sufficient. Real-time supervision must allow visibility into what an agent is doing, what it triggers downstream, and where the trail stops, as quarterly reports are not enough to track the chain of actions. Supervision alone only describes the past: a preventive application is also needed to block a non-compliant call before it executes. A dashboard that reports a violation after the fact is monitoring, not governance. Companies that take responsibility for agents seriously combine supervision and application.

Why Complexity Explodes Between Agents

Companies do not deploy an isolated agent, but fleets of AI agents that call APIs, interact with each other, and access applications that were not designed for machine decision-makers. Adding a second agent already creates a new potential connection, and adding a tenth can generate dozens of interaction paths. Complexity grows with the number of paths between agents, not just with the number of agents, and no one is tasked with mapping these links. Treating governance as a series of one-off approvals is insufficient in the face of a complexity that spans the entire chain of interactions.

Permission Creep and the Gap Between Process and Reality

An agent designed to summarize tickets may receive broad API access because defining permissions precisely would have taken more time. Six months later, this agent may find itself with access to the payment system without any explicit validation being identifiable. More broadly, governance has not reflected how interconnected and cascading agents actually operate, their proliferation outpacing the speed of the processes intended to control them.

Production Rollout and Deployment Pace

Organizations are moving quickly to avoid falling behind, knowing that a slowdown comes at a cost. However, they hit a wall of complexity when betting on agentic AI. Those that manage to overcome this wall have built enough visibility and accountability to maintain control. Complexity is not a reason to slow down: the goal is to achieve a human-agent harmony where scale and responsibility progress together. The risk does not come from a single agent conforming to its design, but from hundreds acting simultaneously in unforeseen combinations, which hinders the transition from pilots to production. Resolving complexity allows autonomy to become an operational goal, while the real danger remains a complex and chaotic system that is difficult to govern.