The fewer agents the better?

getty

How many capabilities can be crammed into an AI agent? A great consolidation may be on the horizon, as it may be far more effective and less costly to add new skillsets into existing agents rather than attempting to deploy fleets of narrow-task agents to accomplish workflows. Even the most technology-savvy leaders are still pondering and probing where the ceiling is in terms of containing agent sprawl and complexity.

This was one of the challenges explored by a panel of industry movers and shakers at the recent Snowflake Summit in San Francisco, which focused on finding the ROI of moving AI and agentic AI from prototyping to production. “You have to be careful in ‘skills’ versus ‘agents’ thinking,” cautioned Sriram Sitaraman, CIO at Synopsys. "Do you want to automate something, or do you want to actually create an agent, which is a different cost structure, usage pattern, and governance and all those things?"

The opportunity – and challenge – is that AI agents can quickly slip beyond the bounds initially set for them, he continued. “Automation is binary. And you can give a human direction. But with agents, it’s a little bit more complicated. Agentic is going to lead you down a path, so you have to be careful."

While the emphasis has been creating and unleashing an agent for every purpose, organizations are finding that adding skills is more productive. “Skills have turned out to be a more agile and smaller unit of currency," said Maddie Want, vice president of data at Fanatics. Examples of more granular skills that can be extracted from existing agents include “codifying a particular piece of knowledge and sharing that across the org. The conversation we have now is does this need to be an agent, or is this just a skill? A lot of the time it’s just a skill.”

Sitaraman’s team started seeing about a year ago that “AI was able to do the work, really good at doing the job of a junior employee – being able to run quick queries and create graphs and create charts and kind of derive insights.” At the time, they started deploying various agents, such as a revenue agent for the finance department that runs reports and a debug agent for the ticketing system to support their data centers.

What the Synopsys team is now moving toward is the idea of a “knowledge agent” that could be deployed “in multiple dimensions – quality, timeliness, and cost-effectiveness." Previously, organizations had to choose two of these dimensions and drop the third, he related. ”By focusing on data, we could move all three metrics more positively," he related.’

AI gets better – not diluted – as it scales. “It doesn’t matter how much data volume you throw at it, because AI is truly a linear scale,” Sitaraman said. "The more data it has, the better decisions it makes.”

This scaling up in AI agent quality was evident at Fanatics, where Want oversees data engineering, data science, and machine learning for the company’s betting and gaming division. “Over time, the degree of investment we had to make in the context layer is decreasing," she related. "And the degree of supervision an agent needs before its able to start autonomously answering questions is decreasing. And our ability to measure the accuracy of the answers is increasing. We can have more confidence in answers without looking."

At the same time, agents were broadening their scope. “Lines are blurring between agents limited in scope, purely analysis agents, and agents that users want to go further and do more with," said Want. "Like blending into operational use cases. because you’ve got the information right there. You can act on that immediately.”

Agents’ roles can be expanded in unexpected directions. “Never underestimate what an agent can do,” Sitaraman continued. "You may have a sales-ops agent, but there’s nothing stopping it from being a sales analyst agent, and a sales-something-else agent.” This is not a process that should develop willy-nilly: frameworks are essential to the intent and context of expanding the scope of agentic-based work.