The AI Race Has a Coordination Problem, Not Just a Safety Problem

30,000 AI agents. More than a billion agent decisions in a single month. And AI already “leading” 26% of the work involved in building better AI inside one of the world’s leading AI labs.

Those aren't predictions about what AI might be capable of in 2030. They are numbers Anthropic reported from inside its own operations in August 2026.

At the same time, 1,386 employees from frontier AI companies have signed a statement arguing that competitive pressure makes it difficult for any company or country to slow AI development on its own. That puts an increasingly uncomfortable question in front of the industry: what happens when everyone sees the need for caution, but nobody wants to be the first to slow down?

Anthropic CEO Dario Amodei has called for the industry to “pace the frontier.” OpenAI CEO Sam Altman has supported elements of that direction, while Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg have pushed back against a coordinated slowdown.

I think this disagreement exposes something bigger than the familiar debate about AI safety versus AI progress.

The AI race has a coordination problem.

AI Is Already Helping Build Better AI

Before talking about slowing AI down, it is worth understanding why this discussion is happening now. Anthropic recently published measurements showing how deeply Claude is being used inside its own AI research and development work.

As of August 2026, Anthropic says Claude “leads” 26% of its measured AI R&D work, while more than 90% involves AI at a collaborative level or higher. Around 30,000 AI agents were also performing research and engineering work at any one time on Anthropic's most-used internal platform.

Anthropic analyzed more than one billion decisions made by those agents during August. About 0.002% of those decisions, roughly one in 47,000, were blocked by its online monitoring system.

But the numbers need context

Those figures sound dramatic, but Anthropic explicitly says Claude is not operating fully autonomously in any measured subset of its AI R&D work. The company also acknowledges limitations in how these measurements are produced, including the use of its own models in parts of the evaluation process.

That distinction matters. Saying “AI is already building itself” would make a much more dramatic headline, but it would go beyond what the evidence currently shows.

A more accurate conclusion is that AI is becoming deeply involved in the process humans use to build the next generation of AI systems.

And that is significant enough on its own.

1,386 AI Employees Are Asking the Same Question

The concern isn't limited to one CEO. In July 2026, 1,386 employees from frontier AI companies had signed the Pacing the Frontier statement, which argues that companies may eventually need mechanisms to slow frontier development temporarily when capabilities advance faster than safety and security measures.

The difficult part is that no company operates in isolation. AI labs compete for researchers, enterprise customers, developers, compute, investment and technological leadership.

Imagine running a frontier AI company while your competitor is training a more capable model, billions are flowing into infrastructure, developers want stronger APIs and customers expect continuous improvement. Even if you genuinely believe more caution is necessary, voluntarily slowing down while everyone else keeps moving becomes extremely difficult.

That isn't only a safety problem. It's a coordination problem.

Sources

https://www.anthropic.com/institute/measuring-pace-of-ai-development

https://www.pacingthefrontier.com/

https://www.anthropic.com/news/accenture-embedded-evaluation

https://www.ft.com/content/2e20c674-3386-46ad-b07f-5cd8a013f67e

https://techcrunch.com/2026/09/20/is-the-ai-industry-really-ready-to-slow-down/