AI Agent Sprawl Is a Work Management Problem

When execution becomes cheap, coordination becomes expensive.

The enterprise AI conversation is moving quickly from copilots to agents. Every major software platform is introducing them. Internal teams are building their own. Specialized agents are appearing for sales, service, finance, operations, product development, and nearly every other function.

The predictable response is to ask how many agents an organization can deploy and how many tasks they can perform.

That is the wrong scorecard.

The real question is whether all that new activity can be turned into coordinated work and completed outcomes.

An organization does not become more effective simply because it has more actors capable of doing things. More actors create more decisions, handoffs, dependencies, exceptions, and opportunities for work to become fragmented. AI agents may lower the cost of execution, but they can dramatically increase the burden of coordination.

That makes agent sprawl a work management problem.

More Agents Do Not Automatically Mean Better Work

Imagine adding hundreds of new employees across an organization without clearly defining their responsibilities, decision rights, reporting relationships, access to information, or measures of success.

No serious leader would expect that to improve performance.

Yet many organizations are approaching AI agents in roughly that way. They are adding capabilities one use case at a time, often inside separate applications and functions, without designing how those agents should participate in the larger system of work.

Each agent may appear productive in isolation:

• One classifies an incoming request.

• Another creates a task.

• Another analyzes the available information.

• Another recommends a decision.

• Another communicates the result.

• Another updates the system of record.

But who owns the outcome across the entire chain? Which agent has the authoritative context? What happens when two agents reach different conclusions? When must a person intervene? How does the organization know the work is actually complete rather than merely processed?

Those are not model-performance questions. They are work management questions.

Cheap Execution Makes Coordination Expensive

For most of organizational history, execution capacity was relatively expensive. Adding another person, team, or vendor required meaningful time and money. That naturally limited how many actors could participate in a workflow.

Agents change that equation. Organizations can add new machine actors much faster than they can add employees. The marginal cost of producing an analysis, generating content, updating a record, or initiating an action continues to fall.

But the cost of coordinating all that activity does not disappear.

It shifts.

Someone still has to determine:

• Which work should be performed

• Which actor should perform it

• What information that actor needs

• How authority should be distributed

• Where work should move next

• Which conditions require escalation

• What controls must be preserved

• How success should be verified

As execution becomes easier to produce, these coordination decisions become more valuable. The bottleneck moves from doing the work to designing how the work should operate.

This is why the next competitive advantage in enterprise AI is unlikely to come from the raw number of agents deployed. It will come from the quality of the work system surrounding them.

The Emerging Agent Sprawl Problem

In a recent HFS Research article examining Asana's agentic AI strategy, Dana Daher argues that enterprises are approaching an agent sprawl problem. Agents may be abundant, but the organizational context, coordination, accountability, governance, and cost control surrounding them will remain scarce.

The warning should sound familiar.

The SaaS era produced application sprawl. Individual teams adopted tools to solve local problems, and organizations later discovered that their data, processes, and work had become fragmented across dozens or hundreds of platforms.

The automation era produced a similar pattern. Teams automated isolated tasks without always improving the end-to-end workflow. In some cases, automation simply moved the bottleneck somewhere else.

Agent sprawl can reproduce both problems at machine speed.

Instead of disconnected applications, organizations may have disconnected machine actors. Instead of isolated automation rules, they may have autonomous agents making decisions and initiating work without a shared architecture.

The result could be enormous amounts of activity with surprisingly little organizational progress.

The Three Work Management Failures Behind Agent Sprawl

The Work Management Institute defines work management as the discipline of clarifying, coordinating, and completing work in a predictable, effective, and sustainable way across an organization.

Agent sprawl puts pressure on all three dimensions.

1. A clarity failure

Agents can only act on the objectives, context, instructions, and boundaries available to them. If the organization has not clarified what the work is meant to accomplish, an agent may complete its assigned action without advancing the intended outcome.

It is possible for every agent in a system to perform its local task successfully while the overall work fails.

2. A coordination failure

Agentic work crosses applications, functions, datasets, and human responsibilities. Work must move between actors without losing context, ownership, or momentum.

If those connections are not designed, agents can duplicate work, make conflicting decisions, trigger unnecessary activity, or create handoffs that no one owns.

3. A completion failure

Agent activity is easy to count. Completed outcomes are harder to define.

An agent may produce a document, update a system, or mark a task complete. But has the customer's problem been resolved? Has the decision been implemented? Has the risk been addressed? Has value actually been delivered?

Organizations must distinguish between machine activity and completed work.

Start With the Work, Not the Agent

The easiest way to create agent sprawl is to begin with a capability and search for somewhere to deploy it.

The more disciplined approach begins with the work.

Before introducing another agent, an organization should be able to answer five questions:

What outcome are we trying to improve?

The unit of analysis should be the outcome, not the isolated task the agent can perform.

How does the work currently move?

Organizations need visibility into the existing workflow, including its triggers, decisions, dependencies, handoffs, delays, and exceptions.

What role should the agent play?

An agent might advise, decide, execute, monitor, coordinate, or escalate. Those are different roles with different levels of authority and risk.

How will humans and agents coordinate?

Human involvement should not be added as a vague promise that someone will remain “in the loop.” The specific points of oversight, approval, intervention, and accountability must be designed.

How will we know the work is complete?

Success must be measured at the level of the intended outcome, not merely by agent usage, task volume, or time saved.

These questions force the conversation away from agent deployment and toward the design of work.

Orchestration Is Not Just a Technical Layer

The software industry often describes the solution to agent sprawl as orchestration. That is partly correct, but orchestration is frequently presented as a technical capability: connecting agents, routing messages, invoking tools, maintaining memory, and managing execution.

Technical orchestration is necessary. It is not sufficient.

An orchestration engine can move information and actions between actors. It cannot, by itself, determine the best division of responsibility, the appropriate decision rights, the intended customer or business outcome, or the organizational consequences of redesigning the work.

Those decisions belong to the architecture of the workflow.

The Work Management Institute develops this argument further in Building Agentic Systems Is Workflow Architecture, which explains why production agentic systems must be designed as systems of work across humans, agents, applications, decisions, and outcomes.

The distinction matters because organizations can technically orchestrate a badly designed workflow. They can connect every component, automate every handoff, and still produce faster confusion.

The Advantage Belongs to Organizations That Can Design Work

Agents will continue to become more capable, more accessible, and easier to deploy. Possessing them will not be a durable advantage.

The more durable capability will be knowing how to integrate agents into the way work is clarified, coordinated, and completed across the organization.

That requires more than AI expertise. It requires an understanding of outcomes, workflows, ownership, dependencies, decision rights, governance, human collaboration, and continuous improvement.

The organizations that develop that capability will be able to add agents without losing control of the work. Those that do not may discover that AI has given them more activity, more complexity, and more coordination debt without producing better outcomes.

The goal is not to build the largest digital workforce.

It is to build a better system of work.