Personal Agent Is the Wrong Product Definition

Instinct is a personal agent, and muse may be a personal agent, but grokbot is not a personal agent and does not want to be one

You can also see this in the bot’s official messaging: most of the examples they share show the bot taking on office work, startup projects, or other long-running tasks. Personal needs are just a subset of the product’s capabilities

The business models also make my point easy to demonstrate. Instinct makes money by taking a cut of transactions users make through the product, meta makes money by using user interaction data to sell precisely targeted ads, and grokbot makes money through subscription fees. The more tokens you burn, the more money they make. So it’s simple: they want you to do everything on it. How many tokens can you burn by booking flights and ordering ubers on it every day?

But people often put them in the same product category, and that is correct: they should be person-like agents. Their ancestor is openclaw. At their core, they package an Agent’s almost unlimited range of capabilities into a single IM conversation, reducing the mental burden on users. Let’s analyze this line of product reasoning in detail

To make an Agent behave like a person, several basic technical problems need to be solved

First, people have memories, while LLMs are stateless and have limited context windows. The first problem that needs to be solved here is agent memory

Precisely because of context window limits, AI products led by chatGPT have used sessions from the outset. But a single session cannot accommodate multiple needs. This is why a growing session list on the left has become the default design of almost every AI product, even though it actually creates a substantial mental burden for users. The only “AI app with nationwide adoption”—Doubao—has clearly taken this problem into account: many people use it every day without even knowing that they can start a new conversation in the upper-right corner

Another problem introduced by the nature of LLMs is reasoning time, which becomes especially noticeable as chatbot products become agents. This directly creates a major pain point in the user experience: every time users send a prompt, they are forced into an annoying waiting period

Because of these two problems, the latest AI products are converging on one form—using a single entry point to handle users’ concurrent requests, hiding intermediate reasoning states, delivering results directly, and encouraging asynchronous conversation. It sounds a little complicated, but plainly put, it’s just IM: the way you communicate with people over IM is the way you communicate with AI in the product

You can assign someone several tasks, and they will quietly handle each request and message you when it is done. You don’t need to see every intermediate step of their thinking. In fact, this product form has been very well received in the market because it solves a critical problem: most AI products, such as office Agents, become increasingly complex, and users simply don’t know how to use them. With this form, users naturally think of the other party as a person who can do anything, making it easy to understand. Instinct takes this form to its extreme: it is even a headless product, with no client of its own. Users interact with the Agent directly in iMessage. muse and grokbot use the same design, which is also a key reason why muse can truly break into the mass market

Next, let’s analyze the core technical designs needed to turn an ordinary ReAct Agent into a grokbot-style person-like agent

First, the agent must respond promptly to user messages. Nobody wants to talk to someone who never replies, right? I’ve analyzed the designs of most products of this kind on the market, and there are roughly two approaches with one shared characteristic. What they have in common is that they must keep the agent loop idle for long periods, rather than executing tasks. The specific approach is to delegate almost all tasks to subagents. In grokbot, this is called the delegate tool, whose description says that all tasks except simple search and question-answering tasks must be delegated through it. Once a delegated task is complete, it calls back into the main loop, activating the loop to communicate with the user

That is what they have in common. The difference lies in what information from the main loop users actually see. A common approach is to show the final assistant text returned by the llm request. But the problem with this approach is that the model won’t “keep quiet”. In a group chat, for example, every message would cause every model to respond, unless you require an @mention to trigger a response. Neither design is particularly elegant. grokbot handles this by telling the model, “the senduser tool is your only voice; the text is just your snippet.” This approach gives group chats and private chats the same structure, making the design more elegant while also making the model less verbose, so it consistently returns only the necessary information