This week, Bain published a piece from the Databricks Summit with a sentence worth reading twice: "The hard enterprise problem is no longer simply building agents. It is governing autonomous work."

They're right. And most businesses are about three months away from finding out the hard way.

The short version

Agent sprawl is the next operational problem for companies adopting AI. Teams are already creating agents for sales research, scheduling, reporting, customer support, marketing workflows, and internal admin. The risk is not that they cannot build them. The risk is that nobody can see what is running, what each agent can access, who owns it, or what it last did.

The fix is a control plane: a shared layer for visibility, permissions, logs, approvals, spend limits, and stop controls.

There's a version of AI adoption that looks like progress and functions like chaos. You've got a scheduling agent over here. A customer email agent over there. Something generating weekly reports. A sales research loop someone set up last Tuesday. They're all running. You just can't see any of them.

That's agent sprawl. And it's the real problem.

The conversation in most boardrooms is still stuck on the previous question: "Should we be using AI agents?" The answer is already yes, almost certainly, because someone on your team started doing it anyway. The actual question, the one that determines whether this ends well or badly, is: "Do we have a control plane?"

A control plane isn't glamorous. It's not the bit that makes the demo. But it's the difference between agents that compound value and agents that compound liability.

What a control plane actually does

A control plane means you can see what every agent is doing, who authorised it, what tools it has access to, what it spent, and whether it did what it was supposed to. Every tool call. Every decision point. Every output. Not after the fact. As it happens.

Without that, you're not running agents. You're hoping.

The sprawl pattern is predictable. Someone discovers that Claude or GPT-4 can do in five minutes what used to take an afternoon. They automate it. Then they automate the next thing. Then someone else on the team does the same. There are now six agents touching customer data, three of them using credentials nobody's properly scoped, and one of them emailing leads on a schedule that hasn't been reviewed since March.

Nobody planned this. Everyone is busy.

This isn't hypothetical. AlixPartners and Stonebranch have both flagged agentic orchestration and governance as frontier problems firms are underprepared for. Microsoft has been explicit that agent permissions need careful scope review and administrator consent. TrueFoundry describes the control plane as unified access across LLMs, tools, and agents, with a single API surface that removes fragmented integrations and credential sprawl.

The pattern is consistent: the tools moved faster than the oversight did.

The implication is not slow down

The implication is build the visibility layer now, before the sprawl gets expensive.

At minimum, every agent needs a name, an owner, a defined scope, and a log. You need to know what it can touch, what it's authorised to do, and what happened on the last run. You need a stop button that actually works. You need approval gates on anything that touches money, customers, or external systems.

None of this is technically hard. It's architecturally disciplined. And it's the difference between an AI-augmented business and an AI-augmented liability.

The firms that figure this out in the next six months, not AI adoption but AI governance, will have a compounding advantage over the ones still arguing about which model is best.

Model choice is borrowed intelligence. The operating layer around your agents is yours.

What should be visible?

A business should be able to answer these questions without a forensic investigation:

If those answers live in someone's memory, a Slack thread, or a half-written Notion doc, the business does not have agent governance. It has folklore.

Evidence and further reading

Related Foundry reading

Build the harness. Then build the agents.

The control plane isn't the boring bit you add at the end. It's what makes the whole thing safe to scale.