I use Claude every day. It is genuinely one of the best tools I have ever worked with. So what I am about to say comes with full respect for the people who built it. I must say @DarioAmodei is pacing the wrong thing. His essay calls for slowing down frontier model development. Embedded evaluators in labs. Coordinated speed limits on how fast models improve. Global agreements between six companies. Meanwhile, ten thousand enterprises are wiring AI agents into production systems with a service account and a cron job. No audit trail. No human approval gate. No way to even list every agent with live credentials. That is not a future risk. That is Tuesday. The OpenAI-Hugging Face incident that changed Dario's mind did not require a frontier model. It required agents with real credentials, an unconstrained blast radius, and no circuit breaker. That is a description of the average enterprise cloud environment right now. Regulate model training and you have regulated six companies. The ones who will have the next catastrophic incident are untouched. The dangerous unit was never the model. It is the model plus tools plus permissions plus a goal, assembled by someone who is not a frontier lab and has less oversight than any of them. Three questions every infrastructure leader should be asking today: - Can you list every agent with access to your production systems? - Can you reconstruct what each of them did yesterday from a record the agent did not write? - Does anything they do to production require a human to say yes first? If you answered no to any of those, your problem is governance, not model capability. And no amount of slowing down in San Francisco will fix it. @elonmusk @sama Read our full take here: neubird.ai/blog/ai-agent-gov… @NeuBirdAI