As tools like Anthropic’s Claude Cowork make it easy for nontechnical employees to spin up task-specific AI “agents,” some large companies are discovering a new flavor of shadow IT: too many bots, doing too many overlapping things, in too many places.
The Wall Street Journal flags “AI agent sprawl” as the problem, fueled partly by commercial platforms and partly by orchestration tools like the open-source OpenClaw. The agents are proliferating faster than corporate governance, which is more or less how every IT department gets its cardio.
The concerns are unglamorous but familiar: security, management overhead, and bills. Magnum Ice Cream CIO Michael Friedlander, whose company owns Ben & Jerry’s, expects duplication — “a lot of people having the same types of agents” — and points to token-driven cost models that may eventually force companies to consolidate and centralize agent use without stomping on employee experimentation.
Gartner, meanwhile, forecasts that within two years the average global Fortune 500 enterprise will run more than 150,000 AI agents. Only 13% of organizations say they have adequate AI agent governance in place. The adoption curve appears to be outrunning the rulebook.
Companies profiled by the Journal are trying to corral the chaos without declaring war on it. FICO’s Mike Trkay says its 3,500 employees are creating “dozens” of new agents daily, prompting governance efforts to avoid conflicting answers from multiple agents aimed at the same problem.
DaVita CIO Madhu Narasimhan says employees have created over 10,000 agents. The kidney-care company blocks “consumer-grade” agent tools from its corporate environment, built an internal platform to pull back on inference and manage token costs when needed, and can increase token spend on its highest-performing agents.
Lyft has rolled out Claude with an IT-approved way to share “skills,” or task-specific instructions, and is building a centralized platform with IT controls for its agents. The concern there is not abstract: Lyft’s Jason Vogrinec noted that too many agents create challenges for “a publicly traded organization with lots of regulatory obligations.”
Anthropic, for its part, says it has added admin features including role-based access, spend controls, usage analytics, audit logging, and curated plug-in libraries.
GitLab CIO Manu Narayan says the company’s existing governance and guardrails are “holding the line,” though he also expects “agent sprawl in the short-term” and says GitLab is “actually OK with that because of the opportunity that AI presents.” Which is one way of saying the bots are multiplying, but nobody wants to be the person holding the flyswatter yet.
Read more at The Wall Street Journal.