TECH Oct 19, 2026 · by Skills Team

AI Agent Sprawl: How to Govern Hundreds of AI Agents

AI agent sprawl is already the normal state of enterprise AI. Learn how identity, permissions, audit trails, and approval gates bring every agent under control.

AI Agent Sprawl: How to Govern Hundreds of AI Agents

The Short Answer

Your company probably has more AI agents working for it today than you think. A coding assistant drafts pull requests, a support bot answers customers at 2 a.m., and a sales agent updates your CRM without anyone opening it. Each one arrived with a good reason, and each one works. The problem is that nobody…

Your company probably has more AI agents working for it today than you think. A coding assistant drafts pull requests, a support bot answers customers at 2 a.m., and a sales agent updates your CRM without anyone opening it. Each one arrived with a good reason, and each one works. The problem is that nobody is keeping count. That is AI agent sprawl: the uncontrolled growth of autonomous AI agents across an organization, each with its own permissions, data access, and habits, and no single place where anyone can see all of them at once.

Sprawl is not a future risk. It is already the normal state of enterprise AI. This guide explains what AI agent sprawl looks like, why traditional IT controls cannot contain it, and how to build a governance model based on identity, least-privilege permissions, audit trails, and approval gates.

What Is AI Agent Sprawl?

AI agent sprawl happens when AI agents multiply faster than the controls meant to manage them. One team adopts an agent to solve a problem, another builds a different agent for a similar job, an employee connects a personal AI tool to company data, and a vendor quietly adds agentic features to software you already pay for. Within months, dozens of agents are reading files, calling APIs, sending messages, and changing records, and no single inventory lists them all.

The result is a workforce you cannot fully see. Agents overlap, duplicate each other, and sometimes work at cross purposes. Worse, each agent is a new doorway into your systems, and agents you cannot list are agents you cannot secure or trust.

How AI Agent Sprawl Differs From SaaS Sprawl

IT leaders have fought sprawl before: a decade ago, the problem was shadow SaaS, employees signing up for cloud tools without approval. AI agent sprawl is harder for three reasons.

First, agents act. A SaaS app mostly stores and displays information, while an AI agent sends emails, moves money, updates databases, and changes configurations. A forgotten SaaS subscription wastes money. A forgotten agent can cause real damage.

Second, agents decide. A traditional integration follows fixed rules, but an agent chooses its next step at runtime, so its behavior can drift beyond what its creator intended.

Third, agents connect to other agents. Modern agents call tools through protocols such as the Model Context Protocol (MCP) and hand off work in multi-agent workflows, so a permission granted to one agent can quietly reach a whole chain downstream. Governance has to cover the network, not just the individual tool.

The Numbers Behind AI Agent Sprawl

Survey data from across the industry shows the scale of the shift.

  • 96% of organizations already use AI agents. OutSystems surveyed nearly 1,900 global IT leaders for its 2026 State of AI Development report: 96% of organizations already use AI agents, and 97% are exploring system-wide agentic AI strategies.
  • 94% are worried about sprawl. In the same report, 94% said they are concerned that AI sprawl is increasing complexity, technical debt, and security risk, yet only a small fraction have built centralized agentic AI governance.
  • 79% name governance as the top scaling barrier. Google Cloud surveyed 1,402 IT leaders for its State of Infrastructure in the Agentic AI Era report, published in July 2026. Security, governance, and MLOps topped the list of obstacles to deploying agentic AI at scale, and 83% said their infrastructure needs upgrades to support production-grade autonomous systems.
  • 150,000 agents per enterprise by 2028. Gartner projections cited by WSO2 suggest the average global Fortune 500 enterprise could have more than 150,000 agents in use by 2028, yet only about 13% of organizations believe they have adequate governance in place.

Together, these numbers tell one story: adoption has happened, governance has not caught up, and the gap is widening fast.

Why AI Agent Sprawl Is a Governance Problem, Not Just an IT Problem

It is tempting to treat agent sprawl as technical housekeeping: find the agents, list them, move on. That misses the real risk. Agents hold delegated authority. When an agent acts, it acts in your company name, with your credentials, on your data, under your legal obligations. If it leaks customer data or approves a fraudulent transaction, the responsibility sits with the business, not with the model.

Governance therefore belongs on the leadership agenda. A governance model answers four questions every executive should be able to ask at any moment: which agents are running, what each one is allowed to do, what each one has actually done, and who approved its authority. If you cannot answer all four today, you have a governance gap, and it grows with every new agent.

The Five Pillars of AI Agent Governance

A practical governance model rests on five pillars.

1. Identity: Every Agent Gets Its Own Name

Every agent needs a unique, verifiable identity, separate from any human user. An agent running under an employee login is invisible in an audit trail: its actions blend with that person's own work. An agent with its own identity produces its own record, which can be reviewed, restricted, or revoked without touching anyone's personal access. Issue identities centrally, tie each to an owner and a purpose, and retire it when the project ends, like a departing employee's account.

2. Permissions: Least Privilege, Enforced at Runtime

An agent should be able to do exactly what its job requires, and nothing more. A summarization agent needs read access to specific documents, not write access to your CRM. This is the principle of least privilege, and it matters more for agents than for people, because agents act at machine speed and at machine scale. Permissions must be enforced at runtime, not just written in a policy document, including across APIs and MCP tool servers. Guardrails such as PII masking and rate limiting add another layer: they limit the damage even if an agent is tricked or compromised.

3. Audit Trails: A Complete Record of What Each Agent Did

Every agent action should be logged in a form a human can read: what the agent did, which data it touched, which tools it called, and when. Audit trails let you investigate incidents, prove compliance, and learn which agents deliver value. Store logs tamper-resistant and separately from the agent itself, so you keep independent evidence when something goes wrong. Centralized observability turns a scattered fleet into a managed workforce.

4. Approval Gates: Humans Decide the High-Stakes Actions

Not every agent action needs human review. Sending a payment, deleting records, publishing content under your brand, or sharing data externally does. Approval gates, sometimes called human-in-the-loop controls, pause the agent before a consequential action and wait for an authorized person to confirm. Place gates where the stakes justify the friction: too few and risky actions slip through, too many and teams bypass the system entirely. A sensible rule: the more irreversible or externally visible the action, the stronger the gate.

5. Lifecycle Management: From Onboarding to Retirement

Agents need the same lifecycle discipline as employees and software: onboarding, review, and retirement. Onboarding means registering the agent, assigning an owner, documenting its purpose and permissions, and testing it in a sandbox before it touches production data. Review means re-checking its permissions and behavior on a schedule, especially after model, tool, or scope changes. Retirement means revoking identity and credentials the day its job ends. Sandboxing a new agent in an isolated environment first lets you observe what it actually does before it can affect real systems.

What Good Governance Tooling Looks Like in 2026

The governance market has moved quickly because the problem demanded it. In September 2026, WSO2 announced the general availability of Agent Manager, an open-source control plane for governing AI agents across any framework, model, or deployment. It combines per-agent, per-environment identity controls, MCP-level governance, and a sandboxed runtime, plus more than 40 built-in guardrails covering PII masking, rate limiting, and more. Its core idea applies even if you never use the product: separate governance from agent logic, so controls survive when teams switch models or frameworks.

Google Cloud has pushed in the same direction, describing agent sprawl as the root challenge of the agentic era and positioning its Agent Gateway as an architectural answer: audit trails for every agent interaction, scoped permissions per agent, and human approval gates before high-stakes actions. The pattern is consistent: governance is moving from scattered point tools into a single control plane that inventories every agent, enforces policy at runtime, and records everything. For a deeper look at the security risks behind these controls, see our guides on AI agents reading browser tabs and Apple's new Mac controls for agent security.

The AI Agent Sprawl Governance Checklist

Work through this checklist in order: visibility first, then control, then ongoing discipline.

StepActionDone When
1Discover every agent, including shadow and vendor-embedded agentsOne inventory, with an owner per agent
2Give every agent its own identity, separate from human loginsNo agent runs under a personal account
3Map each agent's permissions and cut anything beyond its jobEvery agent has documented least-privilege access
4Turn on centralized audit logging for all agent actionsAny agent's last 30 days, reconstructed in minutes
5Add approval gates for spending, sending, publishing, and deletingNo consequential action happens without review
6Govern MCP connections and third-party tool serversEvery callable tool is approved and logged
7Sandbox new and updated agents before productionNo agent reaches production untested
8Schedule quarterly permission and behavior reviewsReviews are calendared with named owners
9Define a retirement process and revoke idle agentsUnused agents lose access within days, not months
10Write an agent policy employees can actually followStaff know how to request, register, and retire an agent

Two habits make the checklist stick. First, make registration easy: if registering an agent takes five minutes and using an unregistered one takes five seconds, people will choose the shortcut. Second, review on a schedule, not after an incident, since permissions drift and agents gain new capabilities.

Building Governance Into Custom Software From the Start

The cheapest time to govern an agent is before it is built. When agents are developed as part of custom enterprise software, identity, permissions, logging, and approval gates can be designed into the architecture instead of bolted on afterwards. Each agent is born with its own identity, its permissions are declared in code, and its actions flow into the same audit logs the rest of the system uses.

FAQ: AI Agent Sprawl and Governance

What is AI agent sprawl?

AI agent sprawl is the uncontrolled growth of autonomous AI agents across an organization without a central inventory, consistent identity and permission controls, or complete audit trails. It happens when teams adopt agents independently, employees connect unapproved tools, and vendors add agentic features, leaving leadership unable to list, secure, or govern every agent.

How many AI agents does a typical enterprise run today?

No single typical number exists: most organizations cannot count their agents accurately. Surveys show 96% of organizations already use AI agents, and projections suggest the average global Fortune 500 enterprise could run more than 150,000 agents by 2028. The practical lesson: the real count is probably higher than the number in your head, which is why discovery is step one.

What is the biggest risk of unmanaged AI agents?

The biggest risk is an agent acting beyond its intended authority with your credentials and data: leaking sensitive information, changing records, or approving transactions nobody reviewed. At machine speed, a single misconfigured agent can repeat a mistake thousands of times before anyone notices.

How does agent governance relate to cybersecurity?

Agent governance is now a core part of cybersecurity. Agents are non-human identities with powerful access, so they need the same discipline as any privileged account: strong authentication, least privilege, monitoring, and rapid revocation. Organizations that fold agents into their existing security program close the gap fastest.

Should we slow down AI adoption until governance is ready?

No. Slowing down usually backfires, because teams adopt agents anyway, just without oversight. Govern in parallel instead: register the agents you have, put guardrails on the riskiest ones first, and make the governed path the easy path. Governance that moves at the speed of adoption is the only kind that survives.

AI agent sprawl is not a sign that your teams did something wrong. It is a sign that AI agents are genuinely useful, and useful tools spread fast. The organizations that win with agents in 2026 will not be the ones with the fewest or the most agents. They will be the ones that can answer, at any moment, which agents are running, what each one is allowed to do, what it has done, and who approved it. Build that visibility, and sprawl becomes what it should have been all along: a managed workforce. If you want help designing a governed agent rollout, contact our team and we will help you put the control plane in place before your next wave of agents arrives.

#AI agents#AI governance#agentic AI#enterprise AI#AI security#MCP