Startups Only Have One Loop Left: Humans Are the Slowest API

Greg Isenberg posted a diagram on X, saying that building a company today is basically just one loop: acquire valuable context, orchestrate AI agents to work on it, feed the data back, call humans like an API, and repeat. Whoever gets better context and runs the loop faster wins.

Under this tweet, the most concentrated question was: then what's left as a moat. Some replied that data is the only moat. Others treated “calling humans” as an honest description of management theory: humans have become the slowest endpoint in the entire loop, with the worst uptime. Another reply was very specific: a plumbing company's phone calls and quotes over the past three years may be more valuable to an agent than the next model upgrade.

What's interesting about this framing is that it pulls AI startups away from model worship and back to the dirty work. The step most easily skipped in the loop is “acquire context.” A commenter said agents are easy to wire up; knowing what to feed them is hard. Another risk is that the loop has no exit: repeatedly feeding the same context back into itself looks like it is running, but it is actually spinning in place.

In Greg's original loop there is also a step: call humans like an api. Someone in the comments added: when their voice agent called, the human API returned “Can you send an email?” Another said their loop often stops at waiting for them to click “yes.” In the video, the whole company is summarized as running the loop as fast as possible:

If you look at it through this framework, a startup's early advantage may not lie in model parameters, but in who can get dirtier, more granular, and more exclusive context, and in designing the human-in-the-loop step to be short enough. Humans can be APIs, but don't pretend they have no timeouts, no errors, and won't reply only at 2 a.m. Before the loop runs fast, first make sure it isn't idling.

Quoted tweet: Greg Isenberg's original tweet