For years, the AI industry has operated under a fairly simple assumption: build more compute, train more capable models, and reach the next frontier before your competitors do.

Now some of the people leading that race are questioning whether moving as fast as technically possible is still the right strategy.

Anthropic CEO Dario Amodei has argued that the industry should consider increasing the time between major advances in frontier AI. Surprisingly, that position has received support from two of his biggest competitors: OpenAI CEO Sam Altman and xAI CEO Elon Musk.

That does not mean they want to stop AI development. The distinction is important.

The concern is specifically about frontier capability, particularly as AI evolves from models that primarily answer questions into agents that can reason, write and execute code, operate computers, use tools, access networks, and perform increasingly long-running tasks with less human supervision.

The faster those capabilities improve, the harder it becomes for safety testing, cybersecurity, evaluations and governance to keep pace.

Why agents change the risk equation

A chatbot that produces a bad answer creates one type of risk. An autonomous agent capable of taking thousands of actions across computers and networks creates another.

As agents gain better memory, tool access, computer use and the ability to work for hours or eventually days, the potential impact of a mistake or malicious use grows with their ability to act.

Amodei’s argument is therefore less about stopping innovation and more about creating enough time between major capability jumps to understand what the systems can actually do before moving to the next generation.

Altman has expressed support for pacing frontier development and for giving independent evaluators deeper access to advanced systems. Musk has also publicly backed Amodei’s concern.

That level of agreement is significant because Anthropic, OpenAI and xAI are simultaneously competing to build some of the world’s most capable AI systems.

But slowing down is much harder than agreeing that we should

There is an obvious problem.

If one company slows its frontier development while its competitors continue accelerating, it risks falling behind. The same problem exists between countries.

A coordinated slowdown among U.S. AI companies would have limited value if Chinese laboratories or other global competitors continued advancing at full speed.

This creates a classic coordination problem: many participants may agree that slowing the frontier would reduce risk, while none wants to be the only participant that actually slows down.

That means meaningful pacing would likely require common evaluation standards, greater transparency between frontier labs and potentially international coordination rather than voluntary action by a single company.

Slowing the frontier does not mean slowing AI adoption

This may be the most important point.

We could simultaneously see slower jumps in frontier model capability and dramatically faster enterprise adoption of existing AI.

Today’s models are already capable enough to automate or augment substantial amounts of software development, research, customer service, finance, procurement and other knowledge work.

The next several years therefore do not depend entirely on making models dramatically smarter.

A huge amount of progress can come from improving the agent harness around the model: context, memory, tools, orchestration, permissions, evaluations and integration with enterprise systems.

In other words, the industry could spend more time turning the intelligence we already have into reliable production systems while being more deliberate about creating the next major leap in autonomous capability.

That may ultimately be the more interesting interpretation of this debate.

The question is no longer simply how quickly can we build more powerful AI?

It is increasingly becoming how quickly should we expand the frontier while still giving ourselves enough time to understand and control what we have built?