The same tools that made AI agents easy to spin up are quietly creating a mess inside Fortune 500 IT departments.
Companies are quietly buried in AI agents, with employees spinning up thousands of bots faster than IT can track them. The cause: tools like Anthropic’s Claude Cowork make agent creation effortless for anyone. Gartner predicts the average global Fortune 500 will run over 150,000 agents within two years.
Picture a kitchen where every cook brings their own sous-chef. Helpful at first. Then there are forty sous-chefs, three of them making the same sauce slightly differently, all of them running up the grocery bill. That is roughly what’s happening inside some of the biggest companies in the world right now, except the sous-chefs are AI agents and the grocery bill is measured in tokens.
What Happened
Businesses fell in love with AI agents, and now they have a new problem: too many of them. The industry has a name for it already, which tells you how fast it arrived. They call it “AI agent sprawl.”
Avid corporate adopters like Lyft, DaVita and GitLab are racing to manage the pileup without killing the enthusiasm that created it. The tricky part is that the thing causing the sprawl is also the thing everyone wants more of.
The Backstory
Sprawl happens because building an agent stopped being an engineering project. Platforms like Anthropic’s Claude Cowork let nontechnical employees create independent bots on their own, and OpenClaw, an open-source tool for orchestrating multiple agents, has only added fuel.
When everyone can build, everyone builds. As Michael Friedlander, CIO of the Americas at Magnum Ice Cream (the company behind Ben & Jerry’s), put it: “Because everybody can do it,” the company expects lots of people ending up with the same types of agents. Those duplicate bots create cybersecurity headaches, management overhead, and rising compute costs.
The numbers back up the worry. Gartner expects the average global Fortune 500 to run over 150,000 AI agents within two years, yet only 13% of organizations believe they have adequate governance for them. That gap is the whole story.
The Plan
The fix isn’t to ban agents. It’s to round them up and put a fence around the pasture.
At FICO (Fair Isaac), 3,500 employees are creating dozens of fresh agents daily, at nearly every tier of the org chart, from individual email-and-brief bots to larger ones that manage data across projects. The company is now building governance to stop multiple agents from spitting out conflicting answers to the same question.
DaVita has gone further. Employees there have created over 10,000 agents, and CIO Madhu Narasimhan keeps “consumer-grade” tools out of the corporate environment entirely. The kidney-care company built an internal platform that can dial inference up or down, throttling token spend when needed and pouring more into the agents that actually perform.
Lyft rolled out Claude to staff, created an IT-approved way to share “skills” (the instruction sets that tell agents how to do specific tasks), and is building a centralized platform with IT controls. For a public company with heavy regulatory obligations, loose agents are a liability.
Anthropic, for its part, has shipped admin features around role-based access, spend controls, usage analytics, audit logging, and curated plug-in libraries.
The Business Model Angle
This is a classic adoption-curve pattern, and founders should recognize it instantly. Whenever a powerful capability gets cheap and self-serve, usage explodes past the company’s ability to govern it. We saw it with cloud spend, with SaaS sprawl, with shadow IT. Agents are just the newest version, moving faster because the barrier to creation is now “type a sentence.”
The strategic lesson cuts two ways. If you’re an operator, the money insight is that the cost of AI is shifting from licenses to consumption. DaVita’s move to throttle inference and fund only the highest-performing agents is essentially portfolio management for compute. Treating tokens like a budget line, not a flat fee, is going to separate disciplined companies from the ones that get a surprise invoice.
If you’re a founder, the opportunity is the governance layer itself. Every sprawl in history has spawned a category of tools to manage it. The companies cleaning up agent chaos, tracking it, costing it, and securing it, are building tomorrow’s enterprise software. There’s a reason a VC like Sapphire Ventures is paying attention to how hard agents are to even locate, since they might be running on a laptop, a server, or anywhere in between.
The Risk
Here’s the honest counterpoint: tidying up too aggressively can strangle the upside. Friedlander himself flagged the tension, wanting to centralize agents without disrupting employee creativity. Over-governance can turn a fast, bottoms-up innovation engine into another approval queue.
GitLab is making the opposite bet. CIO Manu Narayan says existing guardrails are “holding the line,” and the company is fine with short-term sprawl because of the opportunity AI represents. That’s a real strategic fork: clamp down now and protect against risk, or tolerate the mess to capture the upside while it’s hot. Pick wrong and you either bleed money on duplicate bots or fall behind competitors who let their people run.
The deeper risk is conflicting outputs. When dozens of agents answer the same question differently, the organization stops trusting any of them. That erodes the entire value proposition faster than any token bill.
Quick Questions
What is AI agent sprawl?
It’s when a company ends up with way more AI agents than it can track or manage, often with multiple bots doing the same job. It creates security, cost, and management headaches, and it’s spreading fast as agents get easier to build.
Why are companies suddenly creating so many AI agents?
Because building one no longer requires coding skills. Self-serve platforms like Claude Cowork let regular employees create agents on their own, so creation happens everywhere, at every level, all at once.
How many AI agents will big companies actually have?
Gartner expects the average global Fortune 500 to run over 150,000 agents within two years. For context, DaVita employees have already built over 10,000, and FICO staff create dozens more every day.
Is having too many AI agents actually a problem?
It can be. Duplicate agents raise compute costs and create security gaps, and conflicting answers erode trust. But some firms, like GitLab, accept short-term sprawl because the upside of AI is worth the temporary mess.
The Bottom Line
Agent sprawl is what success looks like before someone organizes it. The winners won’t be the companies that move slowest or fastest, but the ones that build the fence while keeping the pasture open. Govern the chaos, meter the tokens, and fund what works. That’s the playbook, and it’s the same one that’s repeated every time a powerful tool becomes cheap enough for everyone to use.
Based on reporting by Belle Lin for The Wall Street Journal. Read the original article here.
