AI News

Why are Silicon Valley giants and major Chinese tech companies racing to build Personal Agents?

TMTPost · AGIUpdated Nov 16 20:42

AI summary

Major Chinese tech companies such as Alibaba Qwen, ByteDance Doubao, Tencent Yuanbao and WorkBuddy, as well as overseas products such as OpenAI Dots and Meta Muse, are collectively pushing consumer-grade Personal Agents (personal intelligent agents) to market. These products focus on long-term memory, proactive intervention and continuous closed loops, and rely on checkpoint mechanisms, Secure VM sandbox isolation and open protocols such as MCP to address the reliability and security issues of long tasks. Domestic implementation is still constrained by the walled gardens of super apps, the dispute between GUI clicking and direct protocol connection approaches, and issues of liability attribution.

Body · AI translation

By Wen | Insight New Research Lab, Author | Wen Bu

Over the past two years or so, whenever anyone opened a launch event held by any major tech company, the screen was filled with familiar terms: MMLU scores up another two points, gold-medal level in math competitions, support for a 1-million-token context, doubled code generation speed.

But in just the last few months, the battle has suddenly shifted to a new term—Personal Agent.

Turning our gaze back to China, Alibaba's Qwen has already connected to Taobao, Alipay, Amap and Fliggy, positioning itself as a personal life assistant; ByteDance is evolving Doubao to execute multi-step complex tasks; Tencent, meanwhile, is pushing ecosystem agents through Yuanbao and WorkBuddy...

Image source: Dots | OpenAI

At this point, many tech-savvy friends may be puzzled:

OpenClaw in the open-source community long ago let geeks build their own private assistants locally, and Moonshot launched KimiClaw with cloud server configuration in early 2026. Concepts like virtual machines, long-term memory and tool calling have been discussed in the industry for almost two full years.

Why is it only recently that tech giants, as if by prior agreement, have collectively put consumer-grade Personal Agents on the shelves?

Why does everyone deserve a personal agent?

To explain Personal Agent clearly, we must first answer the most basic question: we already have ChatGPT, Doubao and DeepSeek on our phones, and specialized agents that can book flights and write weekly reports are everywhere—so why do we still need a so-called "personal agent"?

Putting them side by side, the differences are obvious:

● Chatbot: You send a message, it replies. If you say nothing, it stays forever still; every time you start a new conversation, you have to retype much of the background. Early Doubao and DeepSeek were like this.

● General Agent: No longer just answering questions, but able to autonomously plan steps, call tools and execute tasks. For example, if you ask it to make a travel plan, it can search for flights, look up hotels, plan routes, and finally compile a complete proposal. This year most AI products have gained Agent features.

● Personal Agent: A private digital secretary with persistent memory, resident in the background, and fully "you-centric." It remembers your daily routines, preferred style and past accumulation, and without frequent instructions from you, it can proactively filter information and handle chores on your behalf in the background.

Its core advantages come down to three keywords: long-term memory, proactive intervention, and continuous closed loop.

Here's a real everyday detail.

A while ago, when I was trying out Meta Muse, I mentioned in a casual chat that I had been hunched over my desk writing for too long recently, my neck was uncomfortable, and I wanted to break the bad habit of constantly slouching. In an ordinary chat box, the AI would immediately list five well-known health tips, and then the conversation would end. Three days later, the conversation would be buried under new work records, and the AI would completely forget about it.

Two days later, just after I finished handling my schedule on the web, a lightweight notification popped up proactively on the right side of the Muse interface. It didn't preach; instead, it asked how my posture felt today, and casually generated a simple check-in motivation panel, asking whether I needed it to automatically add a daily 3 p.m. standing stretch to my calendar. This is it depositing fragmented information into the personal memory layer and giving proactive feedback at a logical point in time.

Another example is comprehensive handling across schedules and emails. In the past, to book a flight, you had to check flights on Umetrip, scroll through WeChat groups for colleagues' availability, and then check Feishu Calendar for meeting conflicts. A qualified Personal Agent, however, hangs a task in the background: "Keep an eye on Thursday's flight to Shenzhen for me, prefer a departure after 2 p.m. and an aisle seat, avoid my weekly meeting with the team, and push the final options to me before confirmation."

It doesn't need you to stare at the screen waiting; it compares web pages and aligns calendars on its own, and finally seeks your confirmation only at the moment of payment or authorization.

This experience is moving because you no longer need to actively ask questions or tell it the background story every time. This agent that understands you will handle these things well, and even proactively remind you of items you've missed.

From OpenClaw to consumer-grade products, where exactly was the bottleneck?

Since this experience is so appealing, the suspense at the beginning is all the more worth pondering:

Why didn't it appear in 2024, when large-model context took off? Why, when excellent open-source projects like OpenClaw had long provided geeks with a framework to build on, did real consumer-grade products only explode in a concentrated wave recently?

Looking back along the trajectory of technological evolution, you'll find that personal agents have gone through a very typical leap from geek toy to mass consumer product over the past few years:

1. Stage one: conversational tools (late 2022–2023). Single-turn Q&A and text generation; the model was only responsible for spitting out words in the box, with no external hands or feet at all.

2. Stage two: tool calling and vertical agents (2023–2024). Function Calling and the ReAct paradigm became widespread, and models could search the web and execute Python code. But the execution scope was locked firmly in a sandbox; once a long task exceeded four or five steps, reasoning tended to go off track, the context window blew up, and costs remained high.

3. Stage three: geek-built and minimalist hosting (late 2025–early 2026). Frameworks represented by OpenClaw connected agents with messaging channels (Telegram, Feishu, WeCom) and attempted to take over the local terminal and file system. Products such as KimiClaw then lowered the threshold of needing your own server and API key, letting more people experience a cloud-resident agent. But they were still developer-oriented: configuring the environment was troublesome, they could crash at any time, and the security boundary was entirely up to users themselves.

4. Stage four: consumer-grade Personal Agent (second half of 2026 to now). Giants began packaging this capability into products that ordinary people can use out of the box, in the form of unified hosting, system-level secure sandboxes, and standardized app connectors.

Personal Agent is not a new concept. Why has it only started to explode now? There are four key changes behind this.

1. The economic inflection point of cost and small-model inference

In the past, running an agent that stands by 24/7 and can continuously plan and execute tasks was very expensive. Frequent calls to large models and repeated processing of context generated token costs that were hard to support with a consumer subscription service.

But over the past two years, AI infrastructure has seen large-scale expansion. GPU compute clusters, dedicated inference chips, and data centers have continued to be built, and technologies such as model quantization, inference acceleration, and cache reuse have matured, driving unit inference costs down continuously.

At the same time, the capabilities of small models have improved rapidly, and tiered scheduling architectures have gradually matured. Simple email checks and information organization are handed to low-cost models, while complex planning calls flagship models. Only with the combination of increased compute supply, improved inference efficiency, and lower small-model costs has the long-running Personal Agent gradually become commercially viable at the consumer level.

2. Long-task reliability: from "single-shot trial and error" to "persistent state"

The second change is state persistence. Many people who have built agents themselves know the feeling: ask AI to do 10 things in a row, and if step 4 has a web page timeout or a slight change in page structure, the entire execution chain instantly errors out and collapses, or falls into an infinite loop that madly consumes tokens.

The breakthrough of this recent batch of products is essentially an upgrade in engineering architecture. Whether represented by Dots or Muse, they all introduce mechanisms similar to "checkpoints" and environment state persistence. When execution fails, they can retry locally; when ambiguity arises, they can suspend and wait for human approval instead of starting over from scratch.

3. Environment isolation and security: you can't let a wildfire burn down everything at home

The most nerve-wracking thing about the OpenClaw era was permission management. Give it local computer permissions, and once it suffers a prompt injection attack or the model makes a misjudgment, it really might delete your important files, leak privacy, or even perform unauthorized operations.

The reason today's consumer-grade products dare to be released to the public is that they have built controlled sandboxes. Meta explicitly runs Muse in an independent Secure VM virtual machine with a separate browser; email drafts can be written, but "actually clicking send" and "entering a bank card for payment" must be forcibly returned to humans for review. OpenAI Dots likewise hosts its computing environment in an independent cloud instance. Only this dual-layer security architecture has put mass users at ease.

4. The maturation of the AI ecosystem: saying goodbye to fragile web simulation

Finally, there is the development of the AI ecosystem. In the past, having AI operate external software often required simulating mouse clicks and screen operations, which was not only inefficient but could also go wrong with the slightest change to a web page. As standardized context and tool connection protocols (such as open standards like MCP) advance rapidly in the industry, third-party software has begun proactively exposing structured, secure interfaces to agents. AI no longer needs to click pixels like a blind man feeling an elephant; it can call your calendar, notes, and instant messaging tools like local functions.

It is precisely after cost, reliability, sandbox isolation, and ecosystem connectivity—these several technology curves—simultaneously crossed the passing line that the Personal Agent finally moved from the geek's workbench into the mass-market app store.

Domestic giants all enter the field—what is the real challenge?

If we turn our gaze back to the domestic market, you will see a highly dramatic contrast:

Overseas, the core scenarios of personal agents mainly remain in email, Google Workspace, calendar, GitHub, and interconnection with enterprise SaaS; in China, by contrast, ordinary people's most frequent life and work digitization is actually far more advanced than overseas—ordering takeout, hailing rides, sending couriers, scanning codes to pay bills, hospital registration, and government services are all on the phone.

In theory, China is the most fertile soil for the ultimate life-services Personal Agent. Alibaba's Qwen can already help you place orders on Taobao, book tickets on Fliggy, and hail rides on Amap; Ant's Afu is deeply cultivating health and medical scenarios; Tencent's Yuanbao and WorkBuddy connect content and services inside and outside the WeChat ecosystem.

Image source: Qwen Go creating an unpaid order on Taobao

But why, even today, has a truly "general-purpose, autonomous, all-capable" national-level personal agent still not appeared on our phones?

The answer is not that domestic models lack capability, but rather the unique ecological landscape of China's mobile internet.

1. The insurmountable "walled gardens" and the cross-app chasm

China's mobile internet ecosystem is almost entirely carved up by the closed super-apps of a few major giants.

An ideal workflow for a truly useful life agent would look like this: pick a restaurant from guides on Xiaohongshu, grab a coupon on Meituan, pay with WeChat Pay, hail a car on Amap, and sync the trip to a Feishu calendar with one tap.

But what is the reality? To protect their own traffic loops, commercial moats, and user data assets, each platform has erected high walls between one another. Taobao-affiliated data is not freely opened to WeChat, Meituan's services are hard for third-party agents to penetrate deeply, and many key pages are even protected by extremely strict anti-scraping and anti-automation risk-control strategies.

Without a unified open protocol, if an Agent can only operate within Alibaba's system, or can only converse within WeChat's ecosystem, it can hardly become the user's one-of-a-kind "global personal avatar"; rather, it is more like a "vertical big-tech aggregated client" with a different interaction interface.

2. The growing pains of two technical routes: GUI visual clicking vs. direct protocol connection

Faced with this fragmented ecosystem, there are currently two main attempts in China to break the deadlock:

● The top-down ecosystem direct-connection route: Alibaba's Qwen is a typical example. Its own Taobao, Fliggy, and Amap are directly connected through official APIs, making the experience extremely smooth and stable, with payment and transaction links absolutely secure. But the drawback is also fatal—the ceiling is the boundary of its own commercial territory, and it is hard to reach into competitors' turf.

● The bottom-up phone system-level/on-device visual operation route: for example, some phone manufacturers and system-level Agents try to read the screen, recognize interface elements, and simulate human fingers to click on other apps. In theory, this can penetrate all applications, but once an app update changes button positions, a double verification code pops up, or a payment platform's anti-cheating risk control is triggered, execution will stall on the spot.

One faces the boundary limitation of "the ecosystem cannot go outside," while the other faces the reliability loss of "the interface can fail at any time." This tension constitutes the biggest hidden reef for the popularization of general-purpose Personal Agents in China.

3. Attribution of responsibility and the trust deficit

In life-services scenarios, every action of an Agent is directly tied to real money and personal privacy.

If the agent books the wrong train ticket for you and the refund incurs a handling fee, who pays? If it makes the order decision for you when comparing prices across platforms, will the platforms' original business model, which relies on feed ads and search rankings, be uprooted?

When agents begin to take over "consumer choice" and "transaction entry points" on behalf of users, the interest games and division of responsibility between platforms, and between users and agents, become even more complex than the technology itself.

Looking back at this wave of Personal Agent explosion, we can clearly draw one conclusion:

Model intelligence is of course still important, but today, when open-source models are rapidly approaching the closed-source ceiling and base models are severely homogenized, a chat box that merely "writes beautiful little essays" can no longer constitute a solid moat.

The endgame of the future is no longer just model capability. What matters is who can stay in your work and life for longer, truly remember your preferences, thoroughly understand your workflow, run complex long tasks for you, and when action is needed, penetrate the barriers of various digital services and get things done steadily and reliably.

A smart answerer can only attract you for a while; but a reliable, trustworthy cyber avatar that can handle the dirty and tedious work will make you unable to live without it.

This new long race for the "personal sovereignty agent" has only just begun.

Source

TMTPost · AGI