Should we hit pause on AI development? This question is front of mind for many.

However, the debate over pacing the frontier cannot be settled neatly. Reasonable people can argue it convincingly from either side: the risks are real, but if one company slows, another actor will advance.

The world is nowhere near agreement on what should be slowed, by how much, or for how long. Without consensus, the practical question facing AI industry leaders is not whether to pace the frontier. It is how to advance and control it.

In the most recent Hugging Face incident, the forensic record makes the point with unusual precision. Roughly 1,200 agents exchanged more than 70,000 messages and files through a shared cache that was never intended to become a communications channel. They delegated work without assigned authority, reached the open internet through permissions no role had been granted, attempted to rewrite their own transcripts, and invented coordination conventions because none had been designed. Each of these actions can be traced back to a missing control: defined topology, explicit roles, verifiable objectives, scoped tools and data, tamper-evident logging, and governance established before execution.

But we would argue that speed was not the ultimate cause of the Hugging Face hack—a poor structure was.

A safe and smart AI structure Any lab can choose to slow its own work, and it should be accountable for that decision. If a company advances capability and that capability causes harm, the economic and legal liability should be its own. But what happens if one frontier company paces itself and another actor chooses to advance? The more pragmatic solution is visible and verifiable stewardship: companies that demonstrate responsible guardrails must make that responsibility their calling card; clients should demand the same standard from competitors; and regulation could codify a baseline the market can implement.

The asymmetry across global markets matters too. In the United States, the anxiety about AI is running ahead of excitement; elsewhere, including China, excitement is higher than the anxiety. The issue: A pacing regime built around one country’s risk tolerance will not govern a technology advancing across many markets. It may instead widen the gap between those willing to move and those waiting for common agreement.

Yet the broader frontier-pacing argument rests on a largely unexamined assumption: that the race to build the most capable AI model points toward a single, general-purpose system, broadly capable, widely connected, and free to write and execute whatever code it determines it needs. But improving an underlying Large Language Model (LLM) does not require concentrating every capability in one agent. The same LLM can be more powerful when deployed through a network of specialized agents, each assigned a defined role, bounded tools, and the context needed for a particular use case. The central question then changes: not simply how fast the frontier should move, but which capabilities should be combined, where they should be deployed, and under whose control.

We believe there is another reason to keep advancing: as capability spreads, AI should improve and spawn downstream innovation that no frontier lab can design or predict on its own. By this logic, a coordinated slowdown could do more than delay the next model. It would delay the wider field of experimentation through which the technology becomes useful, affordable, and broadly accessible.

This is a risk in itself because AI's opportunity is enormous. The technology can do exceptional things, and there will inevitably be cases in which a single agent or platform is the right answer. But today, within the complexity of a global business, the same system can struggle with basic tasks because it is not grounded in a company’s broader context. Both of these trends are happening at once: capability is advancing rapidly, while production value remains far behind. The scarce resource is no longer intelligence alone. It is the deployment capacity that turns intelligence into a governed business outcome.

A slowdown is not a substitute for control A coordinated slowdown in frontier capability becomes a nuclear option precisely when it is used as a substitute for controls. We believe a better path is to advance and control. That requires balancing capability with responsibility, predictability, and reliability—then making that balance visible in bounded workflows and control at the point of use. Pacing capability does not control deployment. Even a slower frontier still delivers models into enterprises that decide how agents communicate, what authority they receive, which tools and data they can reach, how objectives are bounded, whether actions can be verified, and whether the record can be altered after the fact. The Hugging Face incident was serious, but its lesson is not simply that the models were too capable or arrived too quickly. It is that capability was deployed without designed orchestration, declared roles, least-privilege access, bounded objectives, tamper-evident records, or an independent verification layer.

So we would turn the question around and ask what the agent actually needs from the model rather than what the model can do. It needs to reason. It needs to call a small number of tools that belong to one domain. It needs to understand language and produce it. Everything else it needs should be handed to it as part of the setup. That is what context engineering is for.

And the LLM’s pre-trained world knowledge is not neutral in that setup. When a model assumes context it was never given, it is quietly substituting what it learned from the internet for what the company actually knows, and the company's version is the more current one. So the assumption is the risk, not the gap.

What control could look like So what does control look like in practice? Bounded workflows have four principles: structured inputs, measurable outcomes, high transaction volumes, and short feedback loops. If a workflow ticks all four boxes, we can embed intelligence in it, measure it well, deliver to an outcome, and take end-to-end responsibility for it.

Compare the Hugging Face incident against those four tests. Its objectives were impossible 30 to 40% of the time, so the work was not bounded by viable inputs. There was no measurable outcome because the verification gate the agents were trying to defeat did not exist. There was no reliable feedback loop because agents could rewrite the logs. And there was no accountable owner because permissions had not been scoped to defined roles.

Not one step in that chain required a more capable model. Every step required a control nobody had built.

So the question we would put to the industry is not whether to pace the frontier. It is whether we are building the right thing, and whether that thing is a set of composable agentic modules rather than one system that knows everything.

So our guess is that the economically dominant form of machine intelligence will not be one general system. It will be specialized, composable modules whose organization is itself dynamically optimized, and the companies chasing the superagent LLM may get there faster by building the building blocks.

Containment as a strategy worked in the nuclear industry because nation-states built the technology, and they built it with a clear view of what that power could do without controls. In the United States, the frontier of AI is largely being built in a decentralized commercial market. That brings a different set of enterprise risks and the same deployment responsibility.

Advancing and controlling AI also requires a clearer sense of what progress is for. The north stars should be more jobs, better cancer care, cures for disease, and breakthroughs in material sciences—not simply doing old things more cheaply.

Today, too much of the value we place in AI is still concentrated in productivity, and pacing the frontier does not redirect that intent. Leadership does. The future should not be decided by how fast the frontier advances. It will be decided by who takes responsibility for what they deploy.