The fundamental difference between Personal Agents and Coding Agents. After following Instinct founder Noah Shinn and analyst Ben Thompson's discussions on Personal Agents, I want to share some of my own thoughts. The direction of the AI industry shifts far too fast. From chatbots to coding agents, then to general-purpose agents, and now to personal agents — in just over a year, the industry has changed direction multiple times. This time, after Muse took off, I felt it very clearly. The focus of everyone's discussion suddenly shifted from model capabilities to cloud computers, Context, long-term memory, proactivity, permissions, and the like. It seems the era of applications is truly arriving. Truly arriving. Of course, how much room for startups remains in this space — I wouldn't dare conclude yet. Meta and OpenAI have already entered the game, and back home, ByteDance, Tencent, and Alibaba are watching like hawks. Many people criticize Personal Agents as nothing more than a hyped-up concept. True, it does look more like a combination of many prior technological innovations that have finally fermented into a new product form. So it's not really hype either. If you read books on the Industrial Revolution, you'll find the technologies of that era were much the same. All innovation is essentially a mix. Here are a few thoughts of my own. 1. The difference between Personal Agents and Coding Agents The biggest difference between a Personal Agent and a Coding Agent is whether it takes action proactively based on Context. For previous Coding Agents, the core logic was still waiting for the user to assign tasks. For example, you'd open Codex and say "fix this bug for me," or have it write a Slides deck. Every task had to start with a user instruction. But a Personal Agent works proactively based on the user's Context. It knows our calendar, emails, location, and spending history, and understands our habits and preferences. So often, we don't even need to tell it explicitly what to do right now — it can figure that out on its own. Here's a simple example. Suppose I have a client visit at 10 a.m. today. The Personal Agent knows about this and knows where the client's company is located. At 9:30, it can check traffic conditions on its own and decide whether taking the subway or a taxi is better today. If a taxi is needed, it can even pre-arrange a ride, then tell me it's time to leave. This feels exactly like what a secretary who knows your work habits very well should be handling. The Instinct product is already doing something similar. If a document needs to be signed by 3 p.m. and it notices you haven't handled it by 2:55, it can even call you directly to remind you. It's no longer just the mode of accepting tasks and then completing them. Personal Agents are starting to have situational awareness. 2. Personal Agents will drive another leap in token consumption From the perspective of token consumption, Personal Agents will mark another important inflection point. Over the past few years, token consumption has gone through several clear inflection points. The first was ChatGPT. People started using models in large quantities to generate answers, articles, and code. The second was o1, when reasoning models started going mainstream. Models began investing more compute in reasoning before answering a question. The third was Coding Agents. A Coding Agent could work continuously for hours — reading files, modifying code, running tests, then modifying again — consuming a huge number of tokens in a single task. Now the fourth inflection point, the Personal Agent, has arrived. Because it's online long-term. In the past, most compute happened when users actively reached out to AI, which gave rise to features like controlling a computer from your phone. But these were all transitional — humans were still in the loop. With Personal Agents, much of the compute may happen in the background. We haven't opened the AI, haven't sent it a message, and it's still understanding Context, checking for changes, and judging whether there's anything that needs attention. AI's working hours are becoming decoupled from humans' online hours. 3. The privacy issue will be solved through technology Of course, Personal Agents didn't emerge out of nowhere. Cloud computers, long-running tasks, proactive reminders, human-like communication — we've seen scattered bits of all of these before. Some of it existed in Coding Agents, and some in OpenClaw. A Personal Agent is really just recombining these elements together. Give an agent a computer that stays online long-term, give it access to plenty of Context, and give it the permissions to execute tasks. Once these elements come together, the new generation of products arrives. Of course, privacy has always been a critical issue that can't be avoided. The more a Personal Agent knows about us, the more capable it becomes. But conversely, the more it knows about us, the more sensitive the data we have to give it. Emails, instant messages, calendars, contacts, documents, bank cards, shopping records — all of this is private. So before Personal Agents truly go mainstream, the issues of privacy and security must be solved. Meta has now built a Secure VM for Muse. Simply put, every user gets an isolated cloud computer, where accounts and data are stored. Their plan is to roll out Muse Confidential VM later this year, at which point the entire virtual machine will be encrypted with keys held only by the user themselves — even Meta won't be able to read the data inside. I think this approach is feasible. In short, privacy issues are being solved thoroughly through technology. 4. The Agent Network follows naturally Once a Personal Agent knows us well enough, an Agent Network follows naturally. I want to write a separate piece on this topic later. Because when an agent knows a person well enough, it's essentially become that person's digital counterpart — or a private secretary who knows us very well. At that point, agents can communicate directly with each other. For example, if I want to schedule a meeting with a friend, the current process is particularly tedious. I ask if he's free on Wednesday, he says he's not, and sends me two times. I realize I'll be on a business trip during one of them, so I send three more times back. If both of us have agents, this becomes much simpler. I tell my agent to find a time this week to meet with him. The two agents can check each other's shared calendars, coordinate a time, and put the meeting directly onto the calendar. So I think Personal Agents may eventually give rise to a new-generation product like WeChat. Agent-to-agent communication can at least eliminate a huge amount of coordination cost in social interactions. Friend gatherings, business meetings, travel planning — much of the information involved doesn't actually need to be communicated back and forth by humans at all. Much of human-to-human communication is, at its core, an exchange of Context, constraints, and intent. Since both parties' agents already have this information, that communication can happen directly between the agents. 5. The next-generation entry point to the internet Personal Agents are very likely to become the total entry point to the next generation of the internet. Think about it — the reason we install so many apps on our phones and use so many SaaS products on our computers is fundamentally to complete specific tasks. This is what product managers call "Jobs to Be Done." Recall Clayton Christensen's classic quote: users hire a product to do a job. That's all. But after Personal Agents, the logic of completing tasks will change completely. In the future, we may just tell the AI what we want to accomplish, and then the AI decides which specific services to call. The UI layer between users and apps will no longer matter. For example, I was just wrestling with a Haidilao reservation for October 6th. It's a family member's birthday, and I needed to open the Haidilao mini-program, find a nearby store, and then select the date and number of people. Extremely tedious. I was thinking this kind of thing could completely be handed off to a Personal Agent in the future. Just tell the agent: "It's a family birthday on the evening of October 6th, help me book a suitable Haidilao reservation." That's it. I believe Haidilao will definitely offer a reservation agent in the future, or at least provide an interface that's convenient for agents to use. Even without a dedicated interface, a Personal Agent could use a cloud computer and operate through Computer Use to open websites or software and complete the task. Some might say: what if a software product simply blocks agents and doesn't offer a corresponding interface? Simple. I think this process will inevitably involve both proactive and passive choices. If you don't open up to agents, you'll be replaced by other new products, and eventually you'll have no choice but to open up. In short, amid the growing pains, all products should embrace agents. I feel that apps, SaaS products, and platforms will all become the suppliers behind agents, and the Personal Agent will become the true total entry point. Because once an agent can complete the vast majority of operations on our behalf, which app or SaaS gets called afterward matters much less. What truly matters is the user's intent. For example, if I'm going on a business trip to Shanghai, I used to have to separately open flight, hotel, and ride-hailing apps myself. In the future, I just need to say I'm going to Shanghai — or even before I open my mouth, the agent will already know this from my calendar, emails, and itinerary. The software behind it is just a tool it calls on. Whoever controls the user's intent captures the most upstream entry point.