George Hotz published a post today about running 69 agents, and the number is almost certainly a joke, but the title is not: “Create value for others and don’t worry about the returns.”
It got 268 points on Hacker News and sparked 112 comments, which tells you something. People responded to it. Whether they agreed or argued, they had something to say about the framing.
The Agent Part
Running many AI agents in parallel is increasingly normal for developers who work on tinygrad-adjacent infrastructure or any sufficiently large codebase. The workflow looks roughly like this:
- Spin up multiple agents against different subtasks
- Let them run concurrently across the codebase
- Collect outputs, review diffs, merge what works
The practical ceiling used to be context window size and API rate limits. Both have improved substantially. Running dozens of agents simultaneously is now a coordination problem more than a compute problem. You spend more time figuring out which tasks decompose cleanly than you do waiting for completions.
Geohot has been doing this kind of large-scale, automated development work with tinygrad for a while. The number 69 is a signal that he’s not taking the framing too seriously, but the underlying practice, running fleets of agents to push a codebase forward, is real.
The Philosophy Part
The more interesting half of the post is the title itself. “Create value for others and don’t worry about the returns” is a simple idea that cuts against a lot of how people talk about building software right now.
The dominant discourse around AI development is shaped by leverage: how much more can I produce, how fast can I ship, what’s my output multiplier. Agents are framed as a productivity tool. That’s not wrong, but it centers the developer’s returns rather than what gets built and who it helps.
Geohot’s version of this, at least as expressed in the title, is closer to what motivates a lot of open source work. You build something because it’s useful, you put it out there, and you don’t structure your effort around what you’ll personally extract from it. tinygrad exists under this kind of logic. comma.ai, to a significant degree, does too.
This is not a naive position. Hotz is commercially minded and has never pretended otherwise. But there’s a distinction between building things that are commercially viable because they’re genuinely useful, and building things optimized primarily for personal returns. The former tends to produce better software.
What 69 Agents Actually Changes
From a practical standpoint, running this many agents simultaneously is a forcing function for writing clear, modular code. Agents fail in interesting ways when tasks aren’t well-specified, when interfaces are implicit, or when side effects are unpredictable. The discipline required to direct a fleet of agents is similar to the discipline required to write code that other developers can work in without constant hand-holding.
The philosophical and the technical converge here. Building for others, whether those others are agents or humans, requires the same kind of clarity: explicit interfaces, documented assumptions, outputs that are easy to evaluate. The returns, as Hotz frames it, aren’t the point. The legibility is.