How to keep an agent focused (and why that’s important)

Agents can’t do everything. In fact, tech leaders at Constellation Research’s AI Forum in New York City on Sept. 24 seemed to prefer the idea of agents doing one thing.

In his work at AI consultancy Soul of the Machine, Kai Lee, VP of AI strategy and transformation, sometimes sees clients deploying agents that have broad responsibilities. An effective agent, he told the conference crowd, must be extremely focused in terms of its objective, what it needs to know, and what is considered “done” and an acceptable output.

“If you don’t set the right objectives up front, [the agent] can retrieve varied responses based on context and data,” Lee said.

The AI forum attendees recommended strategies to narrow an agent’s scope—and shared why the focused effort pays off.

Please be more specific. As anyone who’s yelled “Get me a human!” into a phone in the past 10 years knows, one place facing lots of AI-powered automation is the call center.

Using an agent to help with a task as broad as “customer support,” however, could mean a variety of precise tasks like searching customer histories, generating outreach, triggering discounts, or recommending product changes.

An effective agent requires clear decision boundaries and success criteria, according to Michael Ni, VP and principal analyst at Constellation Research, who spoke with us a day after the presentation. Some tasks, if not properly defined, could have an adverse effect.

“If you tell a service agent you want to minimize handling time, it may rush customers off the call, right?” Ni said.

In that instance, Ni suggested, a more specific choice—with clearly approved actions and KPIs—could be: For customers entering a renewal window with a churn score above a set level, the agent decides whether they qualify for one of three approved retention offers—say, a discount.

The agent, in this theoretical scenario, would factor data like service history, account value, and contract terms into its decisionmaking and escalation parameters.

“What you want is lots of small, little agents that have a very well-bound decision,” Ni told us. “It’s not like I want to replace the call center agent. I want to replace the fact that I have a decision for a retention offer, and I know how to have a flexibility to answer that question very well.”

While Lee, during his panel, noted how focused agents offer focused answers, Ni mentioned other benefits of narrowly targeted agents, including savings on tokens, faster decision speeds, and easier auditability.

Clear and dear. In an IT Brew survey conducted in Q4 2025, more than half of IT pro respondents (54%) cited “lack of clear strategy or defined use case” as their primary AI adoption challenge.

Ni, during our conversation, defined a strategy for making an agent more specific:

• Find a specific job to be done—and all the tasks involved.
• Find a decision that, if done with AI assistance, could relieve a human of some work, or allow that human to make a better, more informed decision.
• Set boundaries. “What can it do? What can’t it do? When does it escalate? I give it only the context and tools that are needed for that job,” Ni said.
• Define “success.” That could include parameters like accuracy, cycle time, conversion, margin, or resolution rates. “You give it specific numbers, and now it knows how to evaluate itself. Then, it can start learning,” Ni told us.

When to consult the rules. Some tasks may not call for AI at all—just deterministic code.

Determining who a caller is and pulling up their information could be a coded rule, for example, requiring no tokens at all. Agents don’t need to take on the full conversation, Ni said, and tasks with structured inputs and predictable “correct” paths are better handled by traditional software or workflow automation.

If your problem can’t accept variations in outputs, stick with probabilistic code, Chris Hallenbeck, SVP and GM of AI & Platform at AI infrastructure company Boomi, told attendees at Constellation’s forum. He prefers an agent that stays on task and doesn’t engage in “massive creativity.”

“I want you to do one task and do it incredibly well over and over again,” he told the audience, referring to AI workflows.