Agent sprawl is a buying problem

Agent sprawl is the uncontrolled growth of AI agents inside an organization. Agents are bought, built or switched on tool by tool, each with its own copy of your context, its own permissions and its own guardrails, until nobody can say how many there are, what they know or what they’re allowed to do.

Gartner’s research puts it more briefly: a state where “too many agents are created and shared too widely.”

It isn’t the same thing as shadow AI. Shadow AI is the agent nobody approved. Most sprawl is approved, one purchase order and one vendor upgrade at a time, by people who each made a sensible call for their own team.

And it’s early. MuleSoft’s 2026 benchmark puts today’s average at 12 agents per organization. Gartner expects a very different number within two years.

At twelve agents, a spreadsheet can still keep up. At 150,000, if half of them still work alone, nobody will be able to say what their company’s agents know, or what they did yesterday.

Because agents are bought the way software was bought: one tool, one team, one budget line at a time. Each purchase makes sense on its own. The sprawl only shows up in the sum.

Most agents don’t even arrive as a purchase. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Every SaaS tool you already pay for is becoming an agent vendor, and its agents come with the renewal.

Each of those agents brings three things with it, and that’s where the trouble starts.

1. Its own context. It reads the vendor’s copy of your customers, products and rules, and that copy drifts from every other one. Why copies drift is the subject of an earlier essay.
2. Its own identity. It acts under a service account, an API key or the login of whoever set it up. When something goes wrong, the log says a person did it.
3. Its own guardrails. Its limits, approvals and audit trail live inside the vendor’s walls and follow that vendor’s idea of safe.

Buy ten agents this way and you haven’t hired ten workers. You’ve bought ten copies of your context, ten permission models and ten audit trails that never meet. The gap was there before the agents: the average enterprise runs 957 applications, and only 27% of them are integrated. Agents inherit that gap, and then they act on it.

Gartner names three: misinformation, oversharing and data loss. In practice they show up as five problems, and each one grows with every agent you add.

1. Contradiction. Two agents act on two versions of the same customer, and each is right by its own data.
2. Oversharing. An agent connected for one task can read everything its connector reaches, and it will use what it reads.
3. No accountability. Without an identity per agent, nobody can say which agent did what or on whose authority, and nobody can switch one off without breaking another.
4. Duplicated cost. The same capability is bought, configured and tuned in five tools. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 over escalating costs, unclear business value or inadequate risk controls. Sprawl produces all three.
5. Shadow agents. Block the official route and people build their own. Microsoft’s Cyber Pulse found that 29% of employees have already used unsanctioned AI agents for work.

Marketing tends to show it first. A martech stack is already a collection of vendors, and each of them now ships its own agents: Salesforce has Agentforce, Adobe has Agent Orchestrator, HubSpot has Breeze. Each one works best inside its own walls.

Picture one customer’s week in that stack.

On Monday the customer complains about a late delivery, and the help desk’s agent logs it. On Tuesday the CRM’s retention agent, which has never seen the ticket, offers them a loyalty discount. On Wednesday the ad platform’s agent retargets them with the same product at full price. On Thursday the email agent sends a “we miss you” message, because in its copy of the data they haven’t opened an email in a month.

Every agent did its job. The customer met four different brands.

The same pattern runs through the rest of the stack.

One customer, three profiles. The CDP, the CRM and the ad platform each resolve identity their own way, so each agent personalizes for a different version of the same person.

A brand voice per tool. Every writing assistant got its own copy of the guidelines, each from a different month.

Consent that doesn’t travel. A customer withdraws consent in one system, and an agent in another keeps acting on the old permission.

Start by counting, then change how agents arrive. Gartner’s six steps begin with governance policies and a central agent inventory, then agent identity and permissions, information governance, monitoring, and a culture of responsible use. Microsoft built Agent 365 as a control plane, with an agent registry and an Entra identity for each agent.

Most companies aren’t there yet.

Do all of it. But notice what an inventory is: a count of the sprawl you already have. It tells you there are 400 agents. It doesn’t give the 401st your context, a proper identity or your guardrails. And Gartner itself warns that blocking agents isn’t a long-term answer, because people route around the block and straight into shadow AI.

Sprawl is a supply problem. It comes from the way agents enter the company, so that’s where it has to be fixed.

An agent factory is one place where a company builds all of its agents, on shared foundations: one context layer every agent reads, one identity model for customers and for the agents themselves, and one set of guardrails every agent runs inside. New agents inherit all three instead of bringing their own.

Shared context. One knowledge layer that maps how the business actually works, built from sourced facts that record when they changed. Every agent reads the same truth, and a correction reaches all of them at once.

Shared identity. One resolved view of each customer, so every agent personalizes for the same person. And an identity for every agent: what it may touch, who owns it, when it retires.

Shared guardrails. One policy, one approval model and one audit trail, applied to every agent, whichever model or vendor runs it. AI agent governance can’t be bolted on afterwards.

Vendor agents don’t have to go. In a factory they become workers: they read the shared context and act inside the shared guardrails instead of bringing their own.

This is also where the economics flip. In a sprawl, each new agent adds another copy of context to maintain, another permission model to audit and another log to check. In a factory, the fiftieth agent is cheaper and safer than the fifth, because everything it needs is already there.

No. The work agents can do is real, and the count will keep climbing whatever any governance team decides. The question isn’t how many agents you run. It’s whether they share one brain or each bring their own.

That’s what we build at House of [MAAD]: the factory underneath, so the agents on top can do real work. Safely.