Box CEO Aaron Levie sees a bright agentic future in the enterprise

AI-backed agents hold the great promise of fundamentally changing how businesses interact with software. Yet people still have a hard time agreeing exactly what an agent is. For simplicity’s sake, I would call it AI-fueled software that can undertake a number of tasks autonomously. 

In the ideal agentic state, they will have agency, meaning they can move across systems, make decisions, overcome obstacles and generally operate without direct supervision.

Leaving aside the lack of a common definition, there are also several key missing or underdeveloped pieces of an agentic system. For this to work right, we will need security, identity, authorization, observability, governance and a host of other capabilities, which are only beginning to be developed. (We saw the release of the AGNTCY project to the Linux Foundation this week that could help with some of these outstanding issues.)

In spite of all this, Box CEO Aaron Levie, who is an ardent advocate for agentic AI, believes that agents will play a central role in enterprise computing in the coming years – even if we have a ways to go to get there.

Levie defines agents as AI assistants that can execute tasks, follow instructions and automate workflows, moving beyond passive information retrieval to active task completion. This is an important distinction from entering a question into a chat interface and waiting for an answer back (that may or may not be accurate)

“It really is going from a world where AI is this kind of read-only experience where you ask a question and it gives you an answer to where it can go off and do real work,” Levie told FastForward.

Just getting started

As Levie sees it, to really understand how this is going to play out, you have to break down the agentic idea into two main parts: architecture and deployment. We’re making progress on the former, how agents are constructed, while deployment is still lagging for now.

“On the architecture of agents, I think we're actually getting increasingly to a mature overall architecture of what agents are, how they're going to work and how they will be constructed. On the actual deployment of agents I think, quite frankly, we're very early,” Levie said. 

And of the deployed agents, the most successful early offerings involve coding tasks. That would be tools like GitHub Copilot, Cursor, Devin from Cognition (the company that bought the remains of coding tool Windsurf earlier this month), among others.

It’s worth noting that even here, SaaStr’s Jason Lemkin recently reported the strange tale of what happened when Replit’s coding agent went haywire and wreaked havoc with his experimental coding project.

Despite that incident, progress outside coding tasks feels much slower than the hype would suggest. But Levie says you have to look at the development of agents as you would any other technological change. It never happens all at once. It’s more of a slow transition. “I think this is a common pattern, which is that those in the tech industry can quickly see the technology curve, and then you can extrapolate out what will happen in the future,” he said. 

He believes that people building agents today are looking out 3, 5 or 10 years as this technology accelerates and anticipating a very bright future. “It becomes obvious that this will completely, radically change everything about how work looks in the future,” Levie said. That will involve building and deploying agents that undertake increasingly sophisticated tasks on our behalf, and as that happens, enterprise software will need to change as well.

Preparing to change enterprise software

Recently, Levie wrote a post on LinkedIn stating that this move to agentic is forcing SaaS vendors to rethink the way users interact with software and that could profoundly change how work flows through companies, something that these firms have to be thinking about as they design the enterprise software of the future. 

“Many of the core design challenges are more about how the user will work with AI Agents to do [a particular] task. This turns the questions into how the user will set up, deploy, orchestrate or provide context to AI Agents to execute work, and then review and incorporate their work after,” he wrote.

In spite of his optimism, he’s also realistic about the current state of things. “Right now, we’re in the messy period where we have to figure some things out,” he said. That includes how you design the architectures, how the agents talk to each other  and co-exist with agents from other vendors and how you secure agents, among other things. These are no small matters, especially to large enterprises, who need a more mature platform before they can start widely deploying agents.

But Levie doesn’t necessarily see this as problematic as you might think. He’s confident that the industry will figure it out, as it has with other major technological shifts. “I think you can be both incredibly optimistic and excited about where this is going, while being pragmatic about the reality that it's going to be years and years of change management and deployment,” he said. 

If agents are truly the future, as Levie believes, it’s essential for enterprises to be looking ahead and preparing for this change, while understanding we could have a long period of transition as we make this shift.

Featured photo courtesy of Box