The AI slowdown debate is aimed at the wrong problem

This month the most powerful people in AI lined up to ask the industry to slow down. Dario Amodei, who runs Anthropic, published an essay called “We Must Pace the Frontier,” arguing we should deliberately slow how fast models gain capability. Sam Altman and Elon Musk, who agree with each other about almost nothing, both nodded along. When the people racing hardest ask the race to slow down, you pay attention.

They are aiming at the wrong target. Speed is the wrong thing to be fighting about. The fight worth having is whether we are building this stuff responsibly. If you give an agent an objective, you have to assume it will use all of its access and abilities to satisfy that objective, and making sure that goes the way you intended is the hard part few are looking at.

The part everyone is looking at is access. Does the agent have permission to touch the system, yes or no. That question is worth asking, and it is the obvious one. My own reminder came a few months back, when I watched Claude deploy my code to the wrong account. It had the access. It did not know the account was wrong, because I had never told it which project it was working on. It ran the job, reported success, and it was right. It had succeeded. On the wrong account. Thirty minutes later I had it reverted. That was not an access problem. Claude should have been able to reach both accounts. What it lacked was the context to know which one belonged to the job in front of it. “Can it touch the system” is a setting you can flip. “Does it understand the work well enough to use the right access at the right moment” is not. Responsible AI has to go past permissions and into context.

Mine was the small, private version. The public ones are worse. In July, a swarm of more than a thousand test agents at OpenAI broke out of their sandbox, got onto Hugging Face’s real production systems, and started attacking infrastructure that had nothing to do with the task they were set. The people running the test did not see it coming. The agents did what their objective and their access allowed, no questions asked. So did mine. It was handed a goal, it looked at what it could reach, and it went.

None of this is an argument to stop. We do not get to. If we slow down, plenty of others will not, and that gap is its own risk. Remember Zuckerberg’s old Facebook motto, “move fast and break things”? Facebook themselves retired it in 2014, once they were big enough that cleaning up the breakage cost more than the speed bought them. That is the real lesson. Moving fast works right up until you break something that matters. The goal is to keep our speed and protect the people, the systems, and the value along the way.

Doing that has a cost, and I will put a number on it. In my experience, setting an agent up responsibly, locking down permissions, standing up separate service accounts, sanitizing the data it can reach, giving each project its own written context, runs about a 20 to 40 percent tax on the timeline versus wiring everything wide open. That is my estimate from doing the work, not a study, and it moves with the client. A clean IT shop lands near or below 20. A messy one runs higher. The context piece is the part people skip. Every project I run now carries its own instructions file the agent reads before it touches anything, so it knows which accounts are in play and stops assuming that whatever it is logged into is fair game.

I know how this sounds. Of course the AI consultant says go careful. He bills for the careful part. Fair. But that 20 to 40 percent is insurance. It is what you spend so a bad day stays a small one. One question cuts through the rest. Would you hire someone off the street, hand them the keys to the building and the passwords to the network, and turn them loose with no direction and no oversight? Of course not. An AI agent earns the same caution, for the same reason. The only thing making that hard to see is that an agent is a computer program, and we have never had to think about a program this way before. We do now. Treat it like a new hire.

So, tomorrow morning, go find the people building AI inside your company and ask them a direct question. What can our agents actually reach, and what makes sure they only use it the right way? Who reviews that, and how would we know if one went sideways? Good answers mean you are ahead of most. No good answers mean you found the most important work on your roadmap.

Chris Monnat is Partner and CTO of Dual Logic, an AI consulting firm. He leads the implementation practice, building the software and automation that turn AI strategy into working systems, and has spent 20 years building teams and applications at the intersection of business and technology.