Matt Shumer@mattshumer_Aug 26
I’m now running agents on four local Macs at once, and each machine’s RAM is completely saturated.
@thsottiaux is 100% right that we need to move past using agents on local machines… when you’re running hundreds of agents at once against a set of increasingly ambitious goals, the bottleneck becomes local RAM and CPU.
Each new model release increases the scope of what I can do, and allows me to effectively use more agents at a time.
Just a few months ago, I was working well on one machine. Then when Fable came out, I had to add two more machines.
Now I’m able to do even more at a time, so I just added another Mac, and I’m thinking about getting another one.
I’ve had to develop a frustratingly complex setup to be able to manage all the agents on all these machines at once. I wouldn’t recommend this to anybody.
This isn’t sustainable.
And I suspect many other folks that are pushing on these models are going to start running into the same problems I am. And soon after, normal users too.
Earlier this year, I built Agent-S (essentially, a precursor to Grok Bot, with a cloud Linux machine for each agent that could scale CPU, RAM, and disk as needed (thx @daytonaio for making this easy!)), and it felt like the future. But the models had to catch up.
Now it’s clear that the models are there, and we need the products to keep pace.
If things keep going the way they’re going, in a year, I will have 50 MacBooks running at capacity in my closet. If this is what actually happens, someone, please, punch me in the face.
We need companies to step up and start building in this direction.
Those that do, and do it right, are going to win massively over the next year.