# What It Takes to Win the Personal Agent Market > I ran all my work through OpenAI's new Dots for a week, and now I am going back to my own setup. Working out why showed me what this market actually asks for. Number: № 034 Published: 2026-10-03 Tags: ai Canonical: https://yihanzhu.com/blog/what-it-takes-to-win-the-personal-agent-market/ Author: Yihan Zhu --- OpenAI held its developer day on September 29 and announced more than twenty things. The one I cared about was Dots, an assistant that stays on all the time and comes with its own computer and browser. I turned it on that afternoon and have been a heavy user since. For the rest of the week, every piece of work I had was handed out through Dots. I should say what it was replacing. I already run my own version of this, which I call my chief of staff. It started from a [daily work brief](https://learn.chatgpt.com/use-cases/daily-work-brief) that OpenAI publishes as an example, and I have added things to it for my job since then. It runs on my machine, next to my code, and the coding work it finds goes to Codex running locally. Dots was supposed to be the same idea done properly, by the company that already had my context. A week later, I am moving my work back to my own setup. Working out why took me past Dots, to a question about the whole market: what does a personal agent need before people will actually use it? ## Dots at work and at home The problems were ordinary. Dots does its work as cloud tasks, on the computer OpenAI gives it. My work lives on my own machine, in local tasks that Codex already handles well, and Dots could not talk to those tasks in any useful way. Then there were permissions. Codex on my machine already has every permission it needs, and Dots kept running into permission problems for the same kind of work. For coding, I would have finished sooner by skipping Dots and giving the task to Codex directly. Compared with my own chief of staff, it was worse on almost everything I asked of it. After a week, my honest verdict is that it is bad. That made me ask what Dots is actually for. I think it is meant to shine at work, and if it is a work tool, it has to be better than handing the job to Codex myself. Right now it is far behind. If it is meant for personal life, the case is weaker still. OpenAI's demo showed Dots booking flights and hotels, the live version failed on stage, and the company [posted it again two days later](https://x.com/OpenAIDevs/status/2105708732323909827) "with better WiFi." Even the working version did not feel close to anyone's ordinary day. The easy reading is that this is a first version and it will get smoothed out. I believe some of that. What I could not shake is that I had picked Dots over everything else before trying any of it, and the reason I did says more about this market than the bugs do. ## Why I reached for Dots first It came down to one fact: OpenAI already knows me. ChatGPT has my conversations and Codex has my code and the way I work. Moving to Dots should have been the smoothest switch available, because the context it needed was already in the same company's hands. The rest of the market works the same way. Meta launched Muse three weeks before DevDay, and it has been [downloaded more than two and a half million times](https://www.pbs.org/newshour/nation/meta-launches-personal-ai-agent-muse-to-help-with-everyday-tasks). People can message it inside WhatsApp, and behind it sit the billions of people already on Facebook and Instagram. Most of them are nothing like me. They do not keep a terminal open, and they will never set up an agent of their own. Muse took off because it is easy to use inside an app they already open, and plenty of people with no technical background already know how to use it. xAI did the same thing with Grok Bot and X. Between them, OpenAI, Meta, and xAI are taking most of this market. The first thing a personal agent needs is a way in. The objection is Instinct, a fourteen-person startup that just [raised a billion dollars at a ten billion dollar valuation](https://siliconangle.com/2026/09/28/everyday-personal-ai-assistant-startup-instinct-raises-1b-at-10b-valuation/), with no app anyone already opens and no larger set of products behind it. Its users are reportedly growing about ten percent a day without any marketing. I have not used it, and I still do not think it lasts. Growth like that is easy to post while a product is invite-only and small, and I heard about Instinct the way I hear about everything here, from the small crowd of people who follow agents closely. OpenClaw and Hermes had that crowd too. The crowd moved from one to the other, and both stalled because they never reached anyone beyond the people who enjoy setting things up. There are not many of those people. As for the valuation, it went up four times in about a month, which mostly tells me how much money is chasing agents right now. ## What I would have to hand over The longer I thought about the way in, the less comfortable it got, because of my own reason for skipping Muse. Muse is the easiest of these to use, and I still have not used it, for one reason: I do not trust Meta with my privacy. For me that outweighs everything else about the product. An agent that books and buys for me has to see my messages and my calendar to do it, and I do not want to expose all of that to Meta. That is the same advantage seen from the other side. The company already standing where you are is the company already holding your context, and an agent only becomes useful once it holds even more. Choosing an agent means choosing whom to hand your life to. I went to Dots first because I had already trusted OpenAI with much of mine, so it barely felt like a choice. The way in and the question of trust turn out to be the same question. There is one way to stop choosing whom to trust, and this week someone raised money for it. [Underdog](https://x.com/0xSigil/status/2106067365733790032) runs its model on your own Mac or iPhone and keeps everything it knows about you on the device, encrypted, with no company server holding it. It claims its model already beats Claude Opus 4.6, the top model of six months ago. I have not tried it, but the direction is the one I want. In *[Own the Memory, Rent the Brain](https://yihanzhu.com/blog/own-the-memory-rent-the-brain/)* I wrote that I could keep my notes at home but still had to send the thinking out to someone else's model, and that I would rather wait for the model to come home. The catch is the first need again. A private agent has the weakest way in of all. You have to hear about it, install it, and decide to move your life into it, which is the path OpenClaw took. ## Agents that have a way in but fall short The companies that make phones already have the biggest way in. Apple shipped [Siri AI](https://www.apple.com/newsroom/2026/09/siri-ai-a-profoundly-more-capable-and-personal-assistant-is-here/) in September, rebuilt as an agent that acts across your apps on every recent iPhone, from a company that has spent years selling privacy. It has the way in, and Apple's privacy story is one people believe. I have used it heavily, and for me it is still unusable. It takes a long time on every request, and what comes back is often wrong. Gemini is the default on Android, so Google holds the same kind of way in and plenty of context. [Gemini 4](https://9to5google.com/2026/09/30/gemini-4-argon-announcement/) posts very high benchmark scores, but the early word from people using it is that it falls short in real work, and I trust that more, because too many benchmarks are hard for the sake of being hard. Google is catching up, and I would not call it the leader yet. Grok Bot looks like the best of all of them to me right now, for work and for personal life both. It [runs on Cursor's cloud agents](https://www.reworked.co/collaboration-productivity/xai-launches-grok-bot-ai-agents-in-beta/), which are mature, and I have watched teams use Cursor's setup for real development work every day. It also connects to X, which is my main platform. What holds it back is the model, which has not been good enough for me, and the cost. Using a stronger outside model through Cursor comes with an extra charge on top, and Cursor's usage limits are less generous than the plans I already pay OpenAI and Anthropic for. Dots belongs in this group too: a way in I already use, a company I already trust, and a product that is not ready. So the third need is the plainest one. Once the agent is in front of you and you have let it see your life, it has to actually do the work, quickly and correctly. Nobody has all three yet. ## Back to my chief of staff My own chief of staff has two of them. I trust it, because it is mine and it runs on my machine. It works, because it lives where my work is, with permissions I granted once. What it lacks is the way in. Almost nobody else would set it up. It is the OpenClaw kind of agent, built by one person who enjoys building it, and it will never win anything outside my own desk. After a week, the work goes back to it. Dots is still there in the Codex desktop app, one entry among the others, but I am no longer sending everything through it. I will look again once it can see the work that runs on my machine. Until then, the best agent I have is one that only I would use.