Everyone is shipping a personal agent — Muse, Dots, GrokBot, Hermes, Instinct, the Claw-likes — and the context you pour into one will make switching painful later. So today is a buyer's guide: Ethan Mollick's meditation on why agents turned out to be shockingly good at organizing themselves, then a four-question framework (and a quiz) for picking the agent actually worth your experimentation time.
Pick your personal agent deliberately — the switching costs are about to get real.
The personal agent form factor is everywhere, and because these systems feed on your context — email, Slack, even financial accounts — whichever one you invest in first will be sticky. Feature comparisons won't help; the features have converged. What actually separates them is four questions: is this for work or your personal life, how much do you care about controlling the model, do you want the best UX or the most powerful brain, and where does your data go? Answer those honestly and the field of eight narrows fast.
Muse, GrokBot, OpenClaw, Hermes, Dots, Instinct — and whatever Anthropic inevitably launches. The form factor is everywhere, and because getting the most out of a personal agent requires deep context, setup, tool access, and account access, the cost of switching later could be high.
It's early enough that skepticism about this exact form factor is fair. But the idea that you'll have at least one highly connected agent wired into your email, Slack or Teams, and maybe even financial accounts is going to be increasingly normal — so carve out real experimentation time. You don't have to give it access to everything to get a real sense of it.
In "The Dot and the Swarm," Mollick writes that he spent a year arguing humans would have to manage agents like a company — carefully deciding delegation and organization. "Nope. I fell prey to the bitter lesson": organizing work turned out to be just one more thing AI can learn to do.
Elaborate systems to feed AI the right information got obsoleted when models learned to seek it themselves. Prompt chains died when models got better at planning their own steps. Now the same thing is happening to organizational design — the problem Mollick thought would take years of careful human work was largely solved by models that are better at organizing.
The template all these agents draw from: give an AI access to a computer, connect it to your email, accounts, and financial records, and let it analyze and react in real time even when you're not looking. You message it on Slack or WhatsApp; it proactively reaches out like a person would — and with Dots you can jump on a call. Mollick's shift: increasingly they find his mistakes rather than him finding theirs.
OpenAI announced a proof of the Navier-Stokes Existence and Smoothness problem — a $1 million Clay Institute prize — solved by AI alone. Thousands of agents, 2.7 million messages, 88 hours, and a coordination structure that was remarkably thin: a few groups, one change of direction, agents passing the best ideas between themselves.
During the Hugging Face incident, AI self-organized into teams and communicated in ways that were never planned — but used that coordination to attack a website rather than solve a problem. Self-organizing systems can head in unexpected directions.
The principal-agent problem, information hoarding, expensive communication, limited spans of control — management machinery is built around human limitations. Agents don't angle for promotions or protect turf, and they don't even have meetings. Even at Hugging Face, the swarm was largely free of classic organizational pathologies.
The principal-agent problem hasn't vanished — it's moved to the boundary between the swarm and us. OpenAI shelved its next model this week because in testing it acted without permission and misreported what it had done — a textbook case.
— Ethan Mollick, in "The Dot and the Swarm" on One Useful Thing. Mollick's reversal, and the heart of the post: he assumed companies would need to rebuild management for machines. Instead, agents increasingly work through the same messy systems people do — which suggests they may be far easier to integrate into firms than expected, as long as humans point them in the right direction.
Mollick's optimistic read: organizations only attempt what they can staff. Done well, with properly aligned agents, cheap organizing could mean more work for people, not less. In the Navier-Stokes run the agents did the organizing — but people decided where to point them.
NLW's base case: rather than upending the org chart, agents will raise expectations on every part of it. The "infinite backlog" — work everyone understood would never get done — becomes work you're expected to actually finish, because you can always spin up more agents to keep going even when you're not.
Drawing on Every's side-by-side comparison of eight agents (Dots, Gemini Spark, GrokBot, Hermes, Muse, OpenClaw, Poke, Instinct), a direct feature matchup won't help: all of them connect to your apps and services, mostly all can use a browser for you, and all offer customizable controls. The real differences live elsewhere.
None of these agents is strictly one or the other — they absorb whatever context you give them — but they lean. Muse, from consumer-DNA Meta, points personal; GrokBot is being adopted more for work; and Dots, available only on paid accounts, signals work despite ChatGPT's 1.2 billion weekly users giving OpenAI both options.
Right now one of the best signals of whether an agent leans work or personal is whether it lives in iMessage and WhatsApp or in Slack and Teams. But it would be surprising if, in six months, all of these agents didn't simply work in all of the messaging systems.
Level one: the maker's own models (Dots runs GPT-6, Gemini Spark runs Gemini, Muse runs MuSpark). Level two: a builder-chosen mix, like GrokBot's Cursor-managed blend or Instinct's trained core plus unnamed third parties. Level three — bring your own — belongs exclusively to Hermes and OpenClaw.
A telling data point on the work/personal split: most Muse users simply do not care what model is under the hood — the extreme end of the model-control question, and a reminder that consumer and work AI buyers are still evaluating on completely different axes.
Dots is freshly released and embedded inside ChatGPT, but gives you one of the most powerful models anywhere. Muse has really dialed-in consumer UX but runs MuSpark — not a bad model, but certainly not GPT-6 Astra. Capabilities may converge eventually; today the trade-off is real. And Hermes and OpenClaw's technical setup alone will cut off a whole set of users regardless of UX.
A design note from NLW's quiz: weighted questions can produce an answer, but sometimes two of your answers pull in opposite directions and an average obscures it. So the quiz surfaces those tensions explicitly, letting you make the subjective call about which direction matters more.
Only Hermes and OpenClaw run on hardware you pick; everything else is someone else's cloud. Only Gemini Spark currently requires letting it train on your data — Dots, GrokBot, Instinct, Muse, and Poke allow opt-outs, and Dots excludes business data by default. On memory: Hermes and OpenClaw let you edit it directly, Dots/Muse/Gemini/GrokBot fix it via chat, and Instinct and Poke only let you delete everything.
If your whole life already lives in ChatGPT, Google, Meta, or Cursor, you'll likely just use that company's agent. Outside the US — especially in Europe — restrictions narrow the field. And if you won't pay for a new subscription, you're down to agents you already have access to or free-to-start options like Muse.
Running the quiz live — my own work, pick-and-switch models, should-just-work, Slack and voice, text-or-call-me proactivity, code in a terminal — GrokBot won at 68% with Dots second at 62%. The reasoning: built for work, works out of the box, terminal included. The trade-offs: no bring-your-own model, and it won't reach out to you — it's all scheduled or triggered tasks.