Gartner's June 2026 projection is worth sitting with: the average Fortune 500 will operate over 150,000 AI agents by 2028. Not 150 agents with ambiguous scope. Not 1,500 agents across a large enterprise. 150,000 task-specific agents embedded in applications, workflows, and business processes across the organisation.
The number sounds extreme. The trajectory that produces it does not. When 40% of enterprise applications embed task-specific agents by end of 2026 - Gartner's separate projection for this year - and those applications scale across a large enterprise's software estate, the agent count compounds rapidly. A single enterprise running 400 applications with even modest agent embedding is already approaching the lower end of that number.
The question that Gartner's projection implies - but that most enterprise AI strategy discussions do not yet directly address - is: what does governance look like at that scale? Not governance in principle, not governance as a set of policies in a document, but governance as operational infrastructure. How does a CIO or CDO maintain visibility, control, and accountability for 150,000 autonomous systems operating simultaneously across every business function?
The answer requires a control plane. And the more interesting question is what happens in the enterprises that arrive at 150,000 agents without one.
What the number actually means operationally
150,000 agents is not an abstraction. It has concrete operational implications that most enterprise governance frameworks are not currently designed to handle.
Consider identity management alone. Each agent has an identity - a set of credentials that grant it access to enterprise systems and data. At 150,000 agents, the credential management problem is larger than the human identity management problem for most enterprises. Human employees number in the thousands or tens of thousands. Agent identities number in the hundreds of thousands - and unlike human employees, they do not have natural lifecycle events (hiring, promotion, departure) that trigger access reviews.
They accumulate permissions over the course of their deployment and may run indefinitely without governance intervention unless the control plane enforces lifecycle management automatically.
Consider audit trail volume. Each agent generates actions - API calls, data accesses, decisions, communications. At 150,000 agents completing tasks continuously, the audit event volume is orders of magnitude larger than traditional enterprise audit logs. The infrastructure that stores, indexes, and makes this volume queryable needs to be designed for this scale. The governance team that reviews it needs analytical tooling, not spreadsheets. Without purpose-built audit infrastructure, the audit trail that regulations require becomes a data storage problem with no practical way to retrieve the specific events that an inquiry will ask about.
Consider the blast radius calculation. A misconfigured or compromised agent at the scale of 150,000 is not an isolated incident. If agents share credential templates, if permission grants cascade across agent families, if a vulnerability in a common agent framework affects multiple deployments simultaneously, the containment problem is not one agent - it is potentially thousands of agents that need to be reviewed, suspended, or remediated simultaneously. Without a control plane that can execute containment at the portfolio level rather than the individual agent level, the incident response timeline is measured in weeks, not hours.
The three governance gaps that compound most dangerously at scale
The governance gaps that are manageable at ten agents and tolerable at one hundred become critical failures at ten thousand and unmanageable at one hundred thousand. Three specific gaps have the most dangerous compounding profiles.
Gap 1: No shared agent identity framework. Most current enterprise AI deployments provision agent credentials ad hoc - each agent deployment creates its own service account, API key, or OAuth grant, managed by the team that deployed the agent using whatever credential management approach that team uses. At small scale, this is inconvenient. At 150,000 agents, it means the enterprise has no unified view of what credentials exist, what they access, and which are still needed. The attack surface from stale or over-privileged agent credentials is enormous and invisible without a unified identity framework.
Gap 2: Inconsistent data access governance. Agents deployed through different frameworks and different SaaS platforms access enterprise data through different channels, with different access controls applied inconsistently or not at all. Some agents access data through properly governed APIs with consent records and audit trails. Others access the same data categories through direct database connections, unmanaged file system access, or vendor-specific integrations with no enterprise visibility. The enterprise's data governance policies apply to some agents and not others - not by design, but because no unified layer enforces them across all access channels.
Gap 3: No portfolio-level observability. Individual agent deployments are monitored by the teams that deployed them - or not monitored at all. There is no enterprise-level view of the agent portfolio's behaviour: which agents are active, which are consuming unusual volumes of data, which are producing outputs that deviate from their defined task scope. Anomalies at the individual agent level are invisible to the enterprise until they produce outcomes - and at 150,000 agents, the anomaly signal is drowned in the noise of normal operations without purpose-built portfolio observability infrastructure.
"Gartner projects AI-governance spend hitting $492M in 2026, with 50% of agent failures tied to weak governance by 2030. The governance gap is not a future problem - it is a present one that is being priced into enterprise AI outcomes right now."
The regulatory dimension at scale
The governance case for a control plane at agent scale is compelling on operational grounds alone.
The regulatory case makes it urgent.
The EU AI Act's high-risk AI system provisions apply to individual AI systems - but at 150,000 agents, the proportion of those agents operating in regulated contexts (employment decisions, financial services, credit assessment, insurance) is significant. Each of those agents, under the Act, requires human oversight mechanisms, explainability, and audit logs. The enterprise cannot manage those requirements agent by agent at scale. It requires a control plane that applies them automatically, at the portfolio level, as a structural feature of how agents are governed rather than as individual compliance efforts per deployment.
The audit trail requirement alone is instructive. The EU AI Act requires logs sufficient to identify situations that may result in risk. At 150,000 agents generating decisions continuously, "logs sufficient to identify situations" is an architectural specification, not a documentation requirement.
The logs need to be structured, queryable, retained for the appropriate period, and indexed in a way that allows the specific event pattern an auditor is looking for to be surfaced without manual review of billions of log entries.
This is infrastructure at data centre scale. It requires design decisions made before the agent portfolio reaches that volume, not after.
What 150,000 well-governed agents looks like
The number 150,000 does not have to be a governance crisis. It can be a competitive capability - if the control plane infrastructure is built before, not after, the portfolio reaches that scale.
A well-governed agent portfolio of 150,000 agents operates differently from an ungoverned one in ways that are measurable and significant.
New agents are deployed into a governance framework that exists - identity registered, permissions scoped, audit logging active - rather than creating governance debt with each deployment.
Anomalies are surfaced at the portfolio level before they produce incidents - an agent accessing unusual data volumes, a set of agents producing correlated output errors, a permission grant that extends beyond task scope - all detected automatically, not discovered through outcomes.
Regulatory inquiries are answered from the control plane's audit infrastructure - the specific agent, the specific action, the specific data, the specific authorisation - in hours rather than weeks of forensic investigation.
Incidents are contained at the data access layer rather than requiring agent-by-agent remediation - revocation at the unified API layer servers access simultaneously across any number of agents that share the same data access pathway.
The difference between the governed and ungoverned 150,000-agent enterprise is not visible at 10 agents. It is not painful at 100 agents. It becomes the defining operational difference at 10,000 agents - and the defining regulatory and reputational difference at 100,000.
The decision that determines the trajectory
Every enterprise AI portfolio starts small and grows. The decision that determines whether 150,000 agents is an asset or a governance crisis is not made at 150,000 agents. It is made now - in the architecture decisions about whether to build agent deployments into a shared governance infrastructure or to allow each team, each business unit, and each SaaS vendor to define their own.
The shared governance infrastructure - the unified identity framework, the governed data access layer, the portfolio-level audit trail, the behaviour monitoring system - is the control plane. It does not need to be built to handle 150,000 agents on day one. It needs to be built at a scale where it can grow with the portfolio rather than being retrofitted onto a portfolio that has already grown beyond the point where retrofitting is practical.
That point is sooner than most enterprises think. The agent portfolio doubles roughly every six months in organisations that are actively building on AI infrastructure. An enterprise with 20 agents today has 160 agents in eighteen months and 1,280 in three years - without any dramatic change in deployment pace. The control plane that needs to govern 1,280 agents is not fundamentally different from the one that governs 150,000. The architecture is the same. The scale is different.
Build it now, for the scale you have, and it grows with you. Wait until you feel the governance pain, and you are retrofitting governance onto a portfolio that was built without it - while simultaneously trying to keep the portfolio operating and answering to regulators who are not interested in your implementation timeline.
150,000 agents is two years away. The governance infrastructure needs to start being built today.
