George Hotz published a post today about running 69 agents simultaneously, and the most interesting part is not the number. The number is a flex, sure, but the actual argument is the title: create value for others and don’t worry about the returns.
This landed on Hacker News with nearly 500 points, and the comments are doing what HN comments usually do with geohot, splitting between people who think he’s a genius and people who think he’s performing genius. Both camps are missing the point.
What he’s actually saying
The context here is agent-driven development. You spin up dozens of AI coding agents working in parallel, pointed at real problems. Some produce useful output. Some produce garbage. The instinct, especially for someone optimizing costs or justifying an experiment to a team, is to immediately ask what the ROI is. How much did this cost per useful commit? What’s the throughput per dollar?
Geohot is arguing that this framing is the wrong one, or at least the secondary one. The primary question should be: did the agents create something useful for someone? If yes, you’ve done something worthwhile. The accounting comes after, not before.
This is a specific and underappreciated point in the current discourse around agentic AI. Most of the conversation is cost-per-task, latency, success rate on benchmarks. These are engineering metrics. They are useful. But they are downstream of a more fundamental question about whether the output matters to anyone.
Why this resonates if you’ve shipped anything
I’ve spent a lot of time building things that run autonomously, bots, processors, scheduled tasks that keep doing work while I’m not watching. The failure mode I run into most often is not technical. It’s motivational. You build something, it runs, and then you spend more time asking whether it’s worth running than you spent building it.
The metrics become a substitute for judgment. If the numbers look good, the thing is good. If the numbers are ambiguous, you stall.
Geohot’s framing cuts through that. Did someone benefit from what you made? Then keep going. Don’t wait for the spreadsheet to give you permission.
The agent angle specifically
Running 69 agents is not something most people can or should do right now. The infrastructure overhead alone is substantial, and the agent tooling ecosystem is still rough. But the principle scales down. If you’re running even one agent, or one automated process, the same question applies.
The trap with AI tooling right now is that it’s easy to measure the wrong things. Token costs are visible. Value delivered to a user is harder to instrument. So people optimize for what they can see and lose track of what they were trying to do in the first place.
Geohot has built enough real things, tinygrad, comma.ai, the various hacks before that, to have some standing when he says the returns will follow if the value is real. That’s not a guarantee. It’s a bet about where to focus your attention while you’re building.
The 69 agents are the setup. The philosophy is the point.