OpenAI’s Brockman lays out five forces shaping the AI agent race

This article summarizes an episode of Alex Kantrowitz’s video series featuring Greg Brockman, president of OpenAI.

Greg Brockman, president and co-founder of OpenAI / Photo credit: Alex Kantrowitz

The next AI bottleneck may be less about clever demos and more about the infrastructure and trust needed to let software take action. Greg Brockman, president and co-founder of OpenAI, frames the race around practical pressures.

These include agents that work across apps, software companies deciding how open to be, permission systems that make delegation safe, scarce compute that supports pricing power, and domain expertise that prevents AI from becoming a commodity.

Persistent agents will reduce the value of app screens

Brockman sees ChatGPT moving from responding to prompts toward doing work. That requires memory, tools, context, and access to browsers, computers, and company software.

As agents handle more tasks, users need to click through fewer apps. He says, “You want almost no interface. You want no product, just the ability to talk to a persistent entity that can go and accomplish goals for you.”

This weakens traditional UI advantage. If the user’s primary relationship moves to an agent, software products compete on usefulness to that agent, not screen time.

The strategic risk is losing direct customer interaction while still carrying workflow responsibility.

Agent access will change how software gets used

That shift creates a hard choice for software companies. They must decide whether AI agents are helpers that increase use or intermediaries that pull control away from the product.

As Brockman puts it, “The idea is: Make your AI specifically good at using [our software].” He adds, “[That software] will be used so much more as a result. If you don’t have AI in there, if you shut it out, you’re going to be declining.”

This creates a new form of platform dependence. Vendors that support agent access may gain volume but surrender some customer ownership. Those that block agents may protect interface control while making their tools less relevant inside automated workflows.

Trust will decide what agents are allowed to do

Once agents can reach software, the next problem is control. Users and companies need to know what an agent can see, change, approve, and send.

Trust sets the limit on delegation. Companies need a way to grant agents different levels of authority depending on the task.

Users need a clear view of what the system knows and what it is allowed to do, with controls that are narrow, visible, and adjustable:

Scarce computing power can protect profits even as prices fall

Expert knowledge will matter more than general AI skill