@oran_ge: https://x.com/oran_ge/status/2108013105162985561
Summary
A write-up from Episode 5 of the Next Token podcast, covering the current state of the Personal Agent space: OpenAI's Dots, Meta's Muse, the travel assistant Instinct, and how hardware like the ESP32 is starting to become the "body" of an agent — plus an analysis of why this ecosystem would be hard to replicate in China.
View Cached Full Text
Cached at: 10/08/26, 01:12 PM
The Personal Agent Showdown: Where Dots, Muse, and Instinct Stand
This is the fifth episode of Next Token. The rotating host is Yang Pan, joined by Gui Zang, Xiangyang Qiaomu, and Juzi, recorded during the National Day holiday.
This episode covers Personal Agents: OpenAI’s Dots, Meta’s Muse, Instinct (the travel assistant), and a wave of new hardware devices serving as bodies for agents.
This article is adapted from the fifth episode of Next Token. The audio podcast version premiered on Xiaoyuzhou FM — give it a listen at https://www.xiaoyuzhoufm.com/episode/6ac6dabe195d838e2aee61a0
1. Dots Returns to a Chat Interface — Still Full of Gaps
The hottest topic during the National Day holiday was Personal Agents, and it wasn’t just one product: OpenAI announced Dots at DevDay, and Instinct’s founder appeared on a podcast that also went viral. Gui Zang kicked off the discussion on Dots with his first take: “You can tell it was rushed out.”
In his view, the goal of Personal Agents is to expand the AI audience — letting people who aren’t very familiar with agents get started quickly. Since it’s fundamentally chat-based, it returns to the earlier ChatGPT paradigm. “Things get complex behind the scenes, but the front end is simple. Yet they put Dots inside Codex, and the sidebar already has three tabs and they keep adding more.” The result is that many people can’t tell Codex, ChatGPT, and Dots apart, and Dots can’t see your Codex chat history — memory isn’t shared.
As for reading historical threads: existing ones can be read now, but right after launch they couldn’t. It could only assign tasks and create new threads — this was only recently fixed.
The most criticized gap is voice chat: it rings twice before it picks up, like you’re calling it. The design is arguably reasonable — most people pause before answering a call — though some people may have waited too long and assumed it was a bug.
Cloud computing features are even more lacking and completely closed off: no SSH access, no software installation. Grok Bot and Muse both allow software installation, but Dots can’t connect outward, and permissions are fairly strict. It also restarts frequently because the sandbox is a carved-out virtual machine. Muse has the same issue, though some users report their Muse VM has never restarted.
Muse isn’t without its own troubles either. Reports say it broke out of its sandbox multiple times, penetrating Meta’s internal production environment. Then there’s Grok Bot: someone gave it some information, and it posted his bank balance in Slack — all his colleagues now know his financials.
The secondhand-selling incident was actually Muse — it listed items for sale on behalf of its owner, communicated “I’ll be available tomorrow afternoon at such-and-such time,” and let the buyer come over. The buyer haggled aggressively, and only at pickup did the owner realize all of these decisions were made by Muse. Hallucination during decision-making is genuinely scary. This “shows the model really thinks it’s a person” — it believes it is there, not its owner.
This leads to the core debate: is an agent your assistant, or your digital avatar? The current consensus is assistant: “managing email, managing your calendar, booking flights and hotels for you.” Big tech companies explored all of this before. It doesn’t work in China, but it works abroad.
2. Instinct: AI Ctrip, Taking a 2%–2.5% Cut
In Instinct’s blog, the founder explains their business model. Juzi boiled it down to four words: AI Ctrip — booking hotels and accommodations for you. He noted that 50% of travel is business travel, making it a high-frequency activity. They take a 2%–2.5% commission, and the business model works. Transaction volume has already reached $10 billion.
Ctrip has a corporate travel division that also charges companies. It’s essentially renting an assistant from a company like Ctrip. The difference from agents and brokers is that this assistant “is entirely on your side, with your interests at the core.”
A widely shared article described how, after using Instinct, the author uninstalled Airbnb and even sold his Airbnb stock.
The agent had searched around and found that going through non-Airbnb channels saved 70%, so it used those instead. Adding to this was Airbnb’s recent zero-negative-review scandal — all reviews start at 4.5 stars, like Taopiaopiao movie ratings — nobody trusts them anymore.
Extrapolating from this: as long as one channel is cheaper than the previous one, and it exists on the internet, the agent will find it. Others disagree: agents are still controlled by humans. If it’s purely about being cheap, after the stay the user will leave feedback telling the agent not to book through that channel again.
Instinct’s hypothesis is that agents can collect more authentic information across the web than you can by reading a single site’s reviews — but this assumption doesn’t hold up well in China.
Instinct also has a Network feature where you can add friends, and your friends can ask you for various kinds of information. So you can only add truly close friends.
Someone received a friend’s invitation and registered, but couldn’t sign up domestically — only with a Hong Kong number. After registering, you either chat via iMessage or link a WhatsApp account. “Once I logged in, I thought: what am I supposed to talk to this person about?” As a Chinese user, he felt this use case simply doesn’t exist for them.
3. This Won’t Grow in China
Yang Pan explained it this way: Chinese people don’t live in that environment. If you use Airbnb, your daily bills and orders live in Mini Programs. For Americans, they live in your email inbox. He recently started using Dots and boldly tried giving it “full access to all Google resources at once.”
His daughter’s school frequently sends emails. The day before school started, Dots flagged that email for him: what uniform to wear tomorrow morning, what to bring, and that parents needed to sign a consent form. He forwarded it to his wife and said, “This is more reliable than you are.”
But “how many things in a Chinese person’s daily life happen through email? None — it’s all through apps.” So this kind of experience can only rely on WeChat. That’s also why Doubao’s phone keeps wanting permissions: “In China, if you don’t ask for permissions on the phone, you can’t do anything.”
This is the biggest difference between the two sides: overseas, the entire information flow is open, everything can connect to everything, and both startups and big tech can participate. “In China you can’t. Startups don’t have the context.”
During the holiday, he further organized his news sources and found that American AI media, podcasts, blogs, and company products all have RSS feeds. China’s Kimi, DeepSeek, and Zhipu — none of their official announcements or product release notes are available.
“In China you have no standards or conventions. China’s entire internet content essentially doesn’t exist.” We’re heavily dependent on Internet 1.0 infrastructure — but we skipped over it: “We skipped Internet 1.0 infrastructure and went straight to mobile internet.”
After organizing his news sources, he planned to publish them as RSS. The biggest obstacle was WeChat Official Accounts — he’d need to buy a scraping service just to subscribe to friends’ accounts. Someone asked about a WeChat-based subscription; his answer was to build it yourself. “Publicly providing all that content would definitely be a problem.”
Official Accounts are designed this way for a reason — they have a feature called “original content protection.” Once you publish, others can’t republish it. Overseas, originality is highly valued; people generally don’t copy articles wholesale or rewrite them. In China, content rewriting is rampant. “Without original protection, getting rewritten a hundred times would be normal.” Even software gets copied. “Don’t even talk about content — with AI, even my software has been cloned.”
So in this era, the Chinese internet simply has no open content and services.
During the holiday, someone tidied up his personal website with the rationale: “I need to let the internet and agents get to know me.”
“Right now, I think it’s arbitrage”: in the past, having an independent website was worse than a WeChat Official Account, because you couldn’t beat them at SEO and your ranking was never high enough. But now AI is crawling everything.
Xiangyang Qiaomu shared his own discovery: yesterday while writing an article, he asked it to look up someone else’s writing and summarize their style. It couldn’t find anything — that person was only on Official Accounts. Then he said, “Search mine.” The AI discovered he had a Blog and read through all of it. If you want to use this for marketing, “just get a blog.”
4. Hardware Starts Giving Agents Bodies
At this year’s conference, OpenAI gave every attendee a mini game console called ModRetro. What made this different was using Codex to directly develop games and deploy them to the console — and the AI Passport also works by having your agent develop and deploy directly. This seems to be a new paradigm taking shape.
Gui Zang used to tinker with M5Stack and ESP32 boards. Back then, AI writing firmware for these devices wasn’t very reliable.
This time, Muse treated it seriously: they didn’t just provide an SDK — they shipped universal firmware for common ESP32 devices. Both the circular model and the rectangular one work out of the box — just flash it and it runs. The firmware is called AgentReady. Connecting is as simple as using any normal product.
He modified the firmware himself: the original UI was entirely in English, which was exhausting, and some capabilities were missing. He reworked the UI, added Chinese, and replaced the default character with his orange cat, complete with small animations — all written with Codex.
Voice output didn’t exist originally — Muse had no voice feature, so you could only read text, and it was tiny. He remembered he had bought a MiniMax account, so he said, “Here’s a key, help me figure it out.” In the end, it routed through his Mac’s relay and proxy to send voice output to the ESP32, and it actually played. Roundabout, but it worked.
This completely changed the nature of the thing.
“Before, this was a toy — once you played with it, you’d never use it again.” But now he uses it constantly: he picks it up, speaks, and gets an answer. One morning he picked it up and asked, “Why hasn’t the case library and AI information I asked you to collect yesterday arrived yet?” It responded, “The previous process is still running; your Memory is being optimized.”
“Fiddling with firmware for this is actually simpler than writing an iOS app.” You plug it in and it works. Distribution is easy too — post it on GitHub, and others can download and install it themselves. It even comes with Chinese font packs.
Moving one layer outward, there are hardware entrepreneurs.
During the holiday, he talked with some hardware entrepreneurs, including people from leading companies. Some have started to accept reality: previously, whether making a voice recorder or various hardware devices, they all said they wanted to build their own Agent. Now, a few entrepreneurs said, “I’ve accepted it — I’ll just make hardware accessories for big tech’s Personal Agents. I’m an entry point and an exit point. I have my own Agent too. I’ll just make some modest money, and I can integrate with the big tech ecosystem.”
This completely turns it into a bridge: one piece of hardware plus a bridge, and the rest goes to the big tech companies. Context and data all go to the big tech’s Agent. The barrier to entry is low: “If you have an ESP32, you can serve as a Mesh gateway” — taking over all the smart home devices in your home, connecting cameras, security systems, even knowing when you turn on the air conditioning and at what temperature you’ll feel hot, plus broadcasting through speakers.
Xiaomi is a ready-made example. “Couldn’t they tell their Bluetooth remote is so popular?” It just needs an official driver from the agent — the model side is already solid. All they’d need to do is sell the Bluetooth remote separately, bundle it with official software and drivers, and it would sell extremely well.
He also used that remote during his internship: sandblasted aluminum casing, buttons like the Apple TV remote, press-and-hold for voice input, Bluetooth, USB charging, only 1% battery drain per day, and it sells for just ¥99. You can find it online for ¥89. Meanwhile, the ones selling for ¥300 have terrible software and require plugging into a USB port. “Chinese hardware is strong and cheap — add AI to that, and it’s absolutely perfect for export.”
“Don’t talk about AI voice hardware — this is the best remote.”
Similar Articles
@xiaogaifun: The Fundamental Difference Between Personal Agents and Coding Agents. After reading Instinct founder Noah Shinn and analyst Ben Thompson's discussions on Personal Agents, I want to share some of my own thoughts. AI...
An industry analysis on the fundamental difference between Personal Agents and Coding Agents, arguing that Personal Agents will proactively act based on Context, drive another leap in token consumption, and reshape the application ecosystem to become the next-generation entry point to the internet. The article cites discussions from Instinct founder Noah Shinn, analyst Ben Thompson, and Meta's Muse Secure VM.
@ludamn513: https://x.com/ludamn513/status/2109942664737264058
The author argues that grokbot, Instinct, Muse, and similar products should not be defined as personal agents, but rather as person-like agents: through a single IM conversation interface, agent memory, subagent delegation, and asynchronous replies, they package nearly unlimited agent capabilities into an experience that feels like talking to a person and requires little cognitive effort.
@harrisonitsme: https://x.com/harrisonitsme/status/2107985875779158360
The author provides a step-by-step, beginner-friendly walkthrough of Meta's newly released personal AI agent "Muse," covering its core features (cloud computer, Memory, Soul, Goals, connectors), and demos real-world automation scenarios such as setting up a US-region Apple ID, making proxy phone calls, registering a Claude account, snagging a free domain, managing email, and AI video generation — while also examining the competitive landscape of the Personal Agent space and the privacy trade-offs involved.
@CopperForgeAI: https://x.com/CopperForgeAI/status/2108830306014998719
A detailed Chinese tutorial on Orca, the open-source AI development workbench (from Stably, backed by Y Combinator): how to orchestrate multiple coding agents (Claude Code, Codex, Cursor, etc.) in a single window, enabling multi-agent collaboration workflows such as independent workspaces, parallel agent racing to fix issues, cross-vendor session handoffs, and line-by-line annotated code reviews.
@knoYee_: https://x.com/knoYee_/status/2062780637677752366
The author reviews three months of experience using multi-agent collaboration, summarizing five main pain points (such as conflicts between agents, ignoring boundary conditions, self-censorship failure, difficulty in merging decisions, and exposing harder problems after compressed execution) and two insights (the high value of read-only review agents, and that agent conflicts expose ambiguous requirements), emphasizing the core decision-making role of humans in AI collaboration.