Grok Bot arrived on 11 August in an early release, and by the time the Moonshots panel got to it, the host had already installed it. "First it was OpenClaw, then there was Hermes, and now there's Grockbot from SpaceXAI," Peter Diamandis said, calling Elon Musk's entry into agent software "the most generally useful consumer AI product that I've seen this year."
The mechanical difference is small to describe and large in practice. Each bot gets its own dedicated cloud computer with a browser, a terminal and the ability to log into your actual applications. You message it the way you would message a colleague rather than typing into a chatbot. In Peter's setup a chief of staff sits on top — his is named Skippy — with specialists beneath it for sales, operations, research and engineering, and whatever other sub-agents you want. The bots run in parallel, message each other, and pull the human in only on judgment calls. The launch announcement describes the same idea from the company's side: persistent teammates that keep working after you stop chatting, hold on to context and pick up your working preferences.
From asking to assigning
Salim Ismail had been playing with it extensively. "I love it," he said. The interface and the ease of use are amazing; what you give up is customizability under the hood. He still thought the trade was worth it, because of what the product changes about the relationship: "you're making a transition from asking an AI to assigning work." A persistent autonomous agent, in his description, is "like a new form of labor" — a completely new category.
He connected it to an idea from his own work on exponential organizations, the attribute he calls staff on demand. Grok Bot, he argued, takes that to its logical extreme: the time spent recruiting staff, and the cost of coordinating them, falls to close to zero. If that holds, he said, the optimal organizational structure changes completely — humans set objectives and leave everything else to the AI.
Eighteen bots, a MacBook and a 5090
Emad Mostaque has 18 Grok Bots running under what he calls the swarm. Through Tailscale, a tool for linking machines into a private network, he has given them control of a MacBook M4 Max, a 5090 graphics card and a range of other computers, plus all of his subscriptions. "So I'm really testing it out," he said.
The cloud computer inside each bot is decent, he said, but you can hand a bot another shell and let it take over an entire machine instead. That is how his sub-bots spend their time. One was installing the new GLM model; another was installing and testing an Alibaba model; a third was optimizing a 27-billion-parameter Alibaba model to run faster on the 5090. By Emad's account, that last one "added 76% to performance at 64K context."
The strangest of his bots does no engineering at all. Atelier has a little team of artists under it and is trying to learn art. Every day it works through its pieces and settles on a main one. "It's not doing very well or it's doing very well, I don't know," Emad said. "I'm not aesthetic enough." He shared one output with the panel: a vinyl record with a piece of hair on it. Someone offered a rule for when to worry — when it produces a banana taped to a wall. The bot had called the work its inner space, Emad said, and the output was getting weird, "but maybe I just don't understand." His verdict on the whole setup was flatter: "it is genuinely useful." Peter expects that usefulness to show up in xAI's revenue; he has raised his own payments to Elon by a few hundred dollars.
A different bill entirely
Dave had not tried Grok Bot. His spending sits on a separate ledger: he had just signed off on $100,000 of swarm agents, with 5,000 instances running that day. He treats the month's theme as abundance of agents rather than one product. All AI interaction to date has been very much one-on-one, he said; now you want a workforce of six, then 50, then a thousand, and he expects people to be talking about 50,000 agents working in parallel within three months.
What that feels like, he said, is trying to manage an organization. "You're like, well, what's everyone doing? I don't know. It's getting really confusing. Are people being productive? I can't tell." His instruction to his team was to go ahead and burn the money but learn quickly, and to work out a reporting structure that can tell whether the agents are doing anything useful. The panel has a phrase for where that leads — the organizational singularity, a company that looks less like an org chart and more like "a continuously orchestrated intelligence network."
"The vaudeville era of agents"
Alex offered the hot take. What is most interesting about Grok Bot, he said, is that it presents like a messaging app — WhatsApp or iMessage, with a pane of the agents in your fleet, conversations you can join, bots talking to each other. "But I don't think that's how it scales. That's completely unscailable."
His reasoning runs through scaling laws, the observed relationships between the resources you spend on AI and what you get back. If growing the size of your agent fleet turns out to be as important as spending more computation at answer time has been for reasoning models, then nobody — not an individual, not an enterprise — is going to hand-manage millions of agents through a chat window. Grafting the interface humans use to manage other humans onto humans managing agents is, in his words, "completely unergonomic." He expects it to be remembered as the vaudeville era of agents: a naive attempt. "The only better manager for agents is other agents, and this doesn't seem to fully internalize that lesson."
Salim agreed, with a caveat about timing. The interface is temporary, something to rotate through while better ones are invented — "but for now, that's a very workable one for coordinating a bunch of agents."
Peter's defense was about adoption rather than architecture. This is what humans are ready to play with, he said; something completely different would get less uptake. He tied it back to the panel's running argument about why AI diffuses slowly: the people one or two layers up the stack from you expect the old thing, so you abstract yourself in a familiar interface. Salim added the exponential-curve version of the same point — the line goes vertical when a technology becomes usable, and what Elon has done at this layer is make AI agents usable to a big set of people. The architecture may not be great, he said; for now it is a powerful entry point.
Peter has invited Alex Finn, who has been making Grok Bot videos, to the Abundance Summit next March, with a condition attached: if Grok Bot is still the hottest thing in March 2027, teach that; if it isn't, teach whatever has replaced it.