New agentic AI tools bring new threat: agent sprawl

You may already have a platform to develop and manage AI agents. You may already have two. Or more. This agent sprawl enables agents to operate in more areas of the enterprise but brings increased complexity, adds to security concerns, and can hurt return on investment.

Over the last year, vendors have been rushing to add agentic AI products to their offerings. Microsoft, AWS, Google, IBM, Salesforce, ServiceNow, Workday, and SAP are already there, while data management software provider Informatica promised this week to add AI agents to its Intelligent Data Management Cloud to automate data processing.

Fears of vendor lock-in

Stiff competition between vendors and fears of vendor lock-in are driving the sprawl, say vendors.

“Enterprises are confused and are holding discussions on which agentic solution to invest in, as nearly all cloud service providers and enterprise SaaS players have some sort of an agentic offering,” said Cam Cross, partner at consulting firm West Monroe.

West Monroe customers are adopting multiple agents as there is no clear vendor winner in the agentic space, and investing in just one platform currently poses a risk of vendor lock-in, said Cross.

And there’s little to discourage enterprises from adding “just one more” agentic AI platform to their mix as vendors are luring them in with free or cheap tools and other offerings, Cross added.

A familiar story

Sprawl, whether agentic AI or something else, is a familiar feature of new product categories in the enterprise technology space. We’ve already witnessed it during the rise of databases, the cloud, generative AI, and large language models (LLMs) — and also with an earlier generation of automation tools.

Likening agent sprawl to the unchecked spread of robotic process automation (RPA) during its heyday, Dion Hinchcliffe, lead of the CIO practice at The Futurum Group, said that RPA too began with small wins, such as automating invoices, onboarding customers etc., but quickly spiraled into a tangle of fragile, overlapping bots with poor governance.

“The same risk applies here: Autonomous agents are like RPA with a brain. They’re smarter and far more adaptable, but without coordination, they’ll collide, duplicate work, and confuse users and systems alike,” Hinchcliffe said.

West Monroe’s Cross sees parallels with the growth in enterprise service bus (ESB) architectures for exchanging data between applications.

“Many technology vendors had their own preferred ESB integration tool, which left enterprises with too many similar tools in-house. Over time, many enterprises had to consolidate these tools to drive down cost,” Cross said.

Whether it’s for RPA, ESB, or agentic AI, having multiple technology platforms for the same task can result in higher operational and support costs, licensing challenges, and having to maintain multiple skillsets.

Can AgentOps stop agent sprawl?

In order to prevent agent sprawl, IT leaders must act decisively before it gets entrenched in their organization.

There are lessons they can learn from previous outbreaks of sprawl, said Paul Chada, co-founder of digital worker platform DoozerAI.

“While database proliferation showed that data governance must precede widespread adoption, cloud sprawl demonstrated that decentralized purchasing without oversight guarantees inefficiency,” said

Hinchcliffe advised enterprises to adopt an AgentOps mindset now. “CIOs should focus on creating an enterprise-wide process for lifecycle management, observability, security, and ROI measurement before agents proliferate,” he said. “Leaders should take cues from DevOps and MLOps: tooling, guardrails, measurement, and central policy layers are key. They should designate a cross-functional team to govern agent adoption, before every department rolls out its own agentic solution.”

A protocol for dealing with agent sprawl

Another way to get on top of the sprawl is for enterprises to focus on interoperability between agents. They can do this by looking for systems that support Model Context Protocol (MCP), a universal, open standard for connecting AI systems to data sources. Another possibility would be to standardize on Google’s A2A protocol for automatic workflows spanning multiple systems.

However, both MCP and A2A are in the early days of development and adoption, and such protocols and interface standards are sure to proliferate. Indeed, the next threat that IT leaders may have to deal with is agentic AI protocol sprawl.