Gartner expects the average global Fortune 500 enterprise to have more than 150,000 AI agents in use by 2028. In 2025, the average was below 15. Manual registration cannot match that pace, which is why AI agent sprawl tends to be discovered rather than planned for.
Less attention goes to what each agent carries. Every agent nobody governs holds access nobody governs, so sprawl is a secrets problem as much as an inventory one. Handling it as both starts with finding the agents already running.
AI Agent Sprawl, Defined
AI agent sprawl is the unchecked growth of AI agents across an organization, built by different teams on different platforms with no shared record of what exists, who owns it, or what it can reach. Individually, the agents may be sanctioned and useful; as a group, they are unaccounted for.
Four conditions produce it:
- Anyone can build one: developers, analysts, and SaaS admins launch agents without a ticket.
- No registry exists, so each agent lives wherever its creator left it.
- Deployment follows no standard path: a notebook here, a CI job there, a SaaS copilot elsewhere.
- Frameworks fragment the estate, since each agent builder describes agents in its own way.
| Dimension | AI agent sprawl | Shadow AI |
|---|---|---|
| The core issue | Too many agents, no central record | AI tools used without approval |
| Who causes it | Approved teams shipping quickly | Individuals working around policy |
| Primary risk | Unowned agents holding live credentials | Company data sent to unvetted services |
| Approved by IT? | Often yes, just untracked | No, by definition |
| First fix | Inventory, owners, expiring access | Discovery, policy, sanctioned alternatives |
Shadow AI is a permission problem. Sprawl also happens inside approved platforms, so a clean shadow AI audit says little about whether the agent estate is under control.
AI Agent Sprawl by the Numbers
- AI agent governance: when Gartner surveyed 360 IT application leaders in 2025, only 13% “strongly agreed that they had the right governance structures in place.”
- Security: Gravitee’s 2026 survey of 750 senior technology leaders found 48% of production agents running unsecured, and 85% of organizations without anyone formally accountable for what agents do.
- Identity ratio: machine identities outnumbered humans 82 to 1 in CyberArk’s 2025 research. The 2026 edition, now published by Palo Alto Networks, puts it at 109 to 1.
Sprawl of Agents Means Sprawl of Secrets
Count the credentials, not the agents. A support agent with read access to a customer database, tickets in two SaaS tools, and an S3 bucket to write to needs four of them. Across a fleet, the useful question becomes how many live secrets the agents hold, and where.
Mostly, those secrets sit where the builder put them: environment variables, MCP config files, notebooks, and CI variables. A GitGuardian scan of public GitHub turned up 24,008 secrets in MCP config files alone, 2,117 of which still worked. That is secret sprawl arriving through a new door, and it folds agents into the wider non-human identity management problem.
| THE COLUMN THAT MATTERS An agent list without a credential column records what exists but not what it can open. Attackers only care about the second part. |
How Do You Find the AI Agents Already Running?
A registry records the agents someone reports. To find the rest, follow the traces agents leave:
- Identity provider app inventories and OAuth grant logs, where any agent that signed in through a consent flow already appears.
- Secrets manager and cloud IAM access logs. Filter for non-human principals whose call patterns match no known deployment.
- Egress and API gateway logs, where scripted user agent strings and call rates no person could produce give agents away.
- MCP server configuration files in repositories and on developer machines, which list the servers an agent can reach and often its keys.
- SaaS admin consoles. Review connected apps, API tokens, and service accounts created in the past twelve months.
Then merge the results on the credential. Two findings that share a key are usually one agent, and a key missing from all your records usually belongs to an unreported agent.
What Should an AI Agent Inventory Record?
Six fields, each useful on its own:
- An owner: one named person, not a team alias.
- Purpose, in a sentence.
- Reach: the systems, data, and tools it can touch.
- The identity it authenticates as.
- Expiry: whether its credential lapses, and when.
- Last used.
The last field does the most work. An agent idle for 60 days is a retirement candidate, and retiring it removes its credentials too, which makes last-used dates the fastest way to shrink an AI agent inventory too large to review.
How Do You Control AI Agent Sprawl Without Blocking Adoption?
Controls that slow teams down make AI agent sprawl worse, because teams under pressure to ship route around them. Three moves work with that pressure:
- Make registration the fast path: a form or API that returns a working identity sooner than a team could wire up its own keys.
- Issue the identity at registration. No team invents its own credential handling, and any agent that can authenticate is on the list by construction.
- Make expiry the default, so access lapses unless renewed and abandoned agents decay instead of piling up.
This is where AI agent governance and credentials meet. Platforms built to govern every AI agent identity, Akeyless among them, issue short-lived access at registration, log each use, and extend the same controls to runtime identity for agents and to machine identity security across the rest of the estate. The deployment guide highlights show how teams roll this out.
FAQs About AI Agent Sprawl
What Is AI Agent Sprawl?
AI agent sprawl is what happens when agents multiply beyond anyone’s records. Each may be approved and useful, but without an owner, a purpose, and known credentials on file, security teams cannot say what the fleet can reach or switch part of it off quickly.
What Is the Difference Between Agent Sprawl and Shadow AI?
Shadow AI is unsanctioned use, such as pasting data into an app IT never approved. Agent sprawl is uncounted use, and it happens on approved platforms as well. A company can close every shadow AI finding and still run hundreds of sanctioned agents with no owner on record.
How Do You Find AI Agents Nobody Registered?
Follow their access rather than their paperwork. OAuth grants, IAM and secrets manager logs, egress traffic, MCP config files, and SaaS connected-app lists all record agents that no registry does, and joining them on the credential each agent uses turns scattered signals into a list.
How Many AI Agents Does a Typical Enterprise Run?
Most cannot say precisely, which is the problem. Gartner’s Fortune 500 figures run from under 15 in 2025 to a projected 150,000 by 2028, but estates vary widely, and only a discovery pass across identity, secrets, and network logs gives your own number.
Who Should Own an AI Agent?
Whoever answers for the outcome the agent produces. Record that person at registration and review ownership whenever someone changes roles or leaves, so no agent keeps running after the last person who could explain it has gone.