Introduction: A “Capable” AI Reminds Me of QQ Pets Twenty Years Ago
When I saw the news that Tencent was opening QQ agents to individuals, I was staring blankly at a test invite code on my screen.
“Raise a genuinely capable AI in 5 minutes”—the headline was like a thorn, piercing precisely into my tech-nostalgia nerve. In 2005, I adopted my first QQ pet in an internet cafe—fed it, played with it, watched it get sick, die, and come back to life. Back then, “virtual pets” were entertainment, social currency, part of youth identity.
Twenty years later, it’s still QQ, still “raising,” but the object has changed from a virtual life needing care into a “digital employee” that can handle Excel, write code, and book flights for us.
This shift itself is more intriguing than any technical parameter.
We stand at a subtle historical juncture: AI’s leap from “tool” to “agent” is happening in the product form most familiar to ordinary people.
Background Analysis: Why QQ of All Things?
To understand this, we need to answer a question first: Tencent has WeChat, WeCom, and the Yuanbao app—why put the debut ground for personal agents on QQ?
This isn’t nostalgia but a carefully calculated product strategy.
1. QQ’s Value as a “Test Bed”
WeChat has 1.3 billion monthly active users, QQ 550 million—behind those numbers are two entirely different ecological niches. WeChat is the social infrastructure for adults, where every change affects the whole; QQ is a digital-native community for Gen-Z, with more tolerance for error and higher tolerance for innovation.
More importantly, QQ’s product DNA is stamped with “virtual identity” and “digital companionship.” From QQ Show to Centimeter Show, from membership tiers to spark badges, this system cultivated users’ habit of emotional connection to “digital entities.” Embedding AI as a “pet” or “assistant” in this context means extremely low psychological switching cost for users.
2. The Technical Tipping Point of Agents
In 2024–2025, LLM capability saw two key breakthroughs:
First, the maturity of multimodal understanding. Early AI assistants could only handle text; now models can simultaneously understand the tables in screenshots, the emotion in voices, and the operation flows in videos. This means AI can truly “see” what users share in QQ and respond.
Second, the improved reliability of tool calling (Function Calling). In the past, having AI call external APIs was like walking a tightrope; now mainstream models have tool-calling accuracy above 95%. This turned “completing real tasks in the chat window” from a technical demo into a deliverable product experience.
Tencent choosing this moment to open personal agents isn’t riding a marketing rhythm but stepping onto the sweet spot of the technology-maturity curve.
3. A Defensive War Over the “Gateway”
A deeper motive hides in the industry landscape. ByteDance’s Doubao, Alibaba’s Tongyi, and Baidu’s Ernie are all fighting for the right to define “AI-native applications.” Tencent’s anxiety is: if users get used to completing all tasks in a standalone AI app, will QQ and WeChat be reduced to “channels that get called”?
Embedding agents into QQ is using an existing scenario to counter incremental competition. You’re going elsewhere to use AI? I’m telling you now: the AI that understands you best is right in the QQ you’ve used for twenty years.
Deep Thinking: From “Chatting” to “Working Together,” What Costs Have We Underestimated?
As a veteran who has lived through multiple paradigm shifts, my excitement about this “embedded AI” is mixed with a professional wariness.
Cost One: The “Colonization” of Attention
The most appealing selling point of QQ agents is “no app switching.” But behind this convenience is a more covert harvesting of attention.
The traditional workflow: I write code in the IDE, open a browser to check docs when stuck, switch to IM when I need to communicate. Every context switch is a cognitive cost, but also a cognitive boundary—it forces me to clarify “what am I doing right now.”
When AI is embedded in the chat window, the boundary disappears. I can discuss requirements with a colleague, @-mention my agent to generate a code snippet, and have it summarize the meeting minutes just now—all at once. Efficiency does rise, but my attention is also cut into finer fragments.
The paradox of efficiency: when we eliminate all “friction,” we may also eliminate all “pauses”—and those pauses are exactly the gaps where deep thinking happens.
It reminds me of an observation at Youzan: we gave merchants extremely sophisticated automation tools, and the result was that many ops colleagues’ work rhythm became “monitor the dashboard—handle exceptions—monitor again.” They became busier, not more at ease.
Cost Two: The “Boiling Frog” of Skill Atrophy
The phrase “raise an AI in 5 minutes” exposes a dangerous metaphor: AI is a pet that needs “raising” but not “understanding.”
When I did technical support at UpYun, I saw too many similar stories. The CDN configuration page became ever “smarter,” with one-click optimization and automatic scheduling; clients really did save trouble. But when genuinely complex network problems arose, fewer and fewer people could troubleshoot independently. Everyone got used to “one click and it’s fine,” and the underlying principles became a black box.
QQ agents’ interaction design is packaging complex task orchestration into simple natural-language commands. The user says “make me a travel plan,” and behind the scenes the AI may call map APIs, hotel-booking interfaces, and calendar-sync services—but the user can’t see this, nor needs to.
In the short term, this is democratization; in the long term, it’s implicit outsourcing of capability. When a generation gets used to commanding AI in natural language to do things that once required programming or professional knowledge, will they still have the ability to “take over the controls when AI errs”?
Cost Three: The “Algorithmization” of Relationship Networks
This is what worries me most, yet is hardest to put into words.
QQ’s core value is social relationships. What happens when everyone’s agent is learning the owner’s conversation style, preferences, and even emotional patterns?
Imagine this scene: your friend’s agent and yours chat first. They discover you both like a certain game and follow a certain blogger, so they proactively suggest “you two masters should grab a meal.” This sounds thoughtful, but the natural growth of relationships is being accelerated by algorithms.
More subtly, as your agent understands you more, will it start “filtering” the information you receive? Just as today’s recommendation algorithms decide what videos we see, tomorrow’s personal agents may decide what messages we see, what invitations we reply to, and what relationships we maintain.
We’re not raising an AI; we’re entrusting our digital persona to a model trained by a commercial company and driven by commercial interests.
Case: An “Imperfect” Agent Experiment
In the second half of last year, I used my spare time for a small experiment: building a personal workflow assistant on an open-source framework, trying to replicate the core experience of a QQ agent while retaining full control.
The tech stack was simple:
- Base model: locally deployed Qwen2.5-72B
- Memory system: long-term memory based on a vector database + a structured event log
- Toolset: calendar, email, and code repository connected via MCP (Model Context Protocol)
- Interaction layer: a minimal chat interface
I set its “persona” not as an assistant but as “a colleague with a very good memory”—it wouldn’t offer suggestions proactively, only provide information when asked; it wouldn’t predict my needs, only record things I explicitly told it.
Three months in, a few interesting findings:
First, “less smart” was actually more reliable. When I disabled the model’s “proactive inference” ability and kept only retrieval based on explicit memory, the hallucination rate dropped from about 15% to under 2%. The cost was that interactions became “clunky,” but the trustworthiness of output rose sharply.
Second, memory needs a “forgetting mechanism.” Early on I kept all conversation records, and soon found the model polluted by outdated information. Later I introduced a simple rule: memories not called for over 90 days go into an “archived” state requiring a specific command to activate. This simulated human memory decay and unexpectedly improved relevance.
Third, the hardest part isn’t technology—it’s “boundary negotiation.” In my interactions with this assistant, the most time-consuming part was clarifying “what it should do and what I should do myself.” This negotiation process is essentially redefining my work identity—which skills are the “core capabilities” I want to keep and which can be outsourced.
This experiment convinced me: the value of a personal agent isn’t in how many things it can do but in how it helps us understand ourselves more clearly.
Viewpoint: We Need “Working Together,” Not “Keeping a Pet”
Back to Tencent’s QQ agent. My attitude toward this product is cautious optimism, but I hold clear criticism of its marketing language.
The metaphor of “raising an AI” is harmful. It implies an unequal relationship: the AI is dependent on us, needs to be cared for, and can be shaped to our will. This framing dodges the real issue—when AI has autonomous execution ability, the human-machine relationship is closer to collaboration than master-servant.
The personal agent I envision should have the following traits:
1. Auditable Autonomy
It’s not “black-box execution” but “white-box negotiation.” When it wants to use my email to send a message, it should show the draft and ask for confirmation; when it wants to access my calendar to schedule a meeting, it should explain the conflict-detection logic. This “visibility” isn’t the enemy of efficiency but the cornerstone of trust.
2. Degradable Capability Design
A good agent should be like a good coach: it helps me complete the current task while ensuring I retain the ability to complete it independently. This means having a “teaching mode”—when I ask it to “write this code for me,” it can ask back, “Do you want to try it yourself first? I can offer some ideas.”
3. Neutrality of Relationships
It shouldn’t “optimize” my social relationships. If I discover two friends share a common interest, it should be my active choice whether to introduce them, not the agent deciding for me. Its role is to enhance my social agency, not replace my social judgment.
The ethical bottom line of technology isn’t what it does but whether it preserves the user’s freedom to “choose not to do.”
Summary and Outlook: Between “Intelligence” and “Wisdom”
The opening of QQ agents marks the personal AI agent’s move from geek toy to mass market. This is a milestone worth celebrating, but also a turning point worth watching.
As a practitioner, I see the maturity of the tech stack, the improvement of engineering capability, and the validation of the productization path. As a user, what I care about is where my data goes, the boundaries of my attention, and the risk of my skill atrophy. As a member of society, what I worry about is: when the daily decisions of a billion people are all embedded in the same AI infrastructure, will diversity and serendipity be systematically compressed?
Advice for ordinary readers:
If you’re an early adopter, consciously observe your interaction patterns with AI. Record which tasks you’ve handed over and which you’ve kept for yourself, and periodically review whether this boundary is reasonable.
If you’re a product manager, resist the temptation of “omniscience.” The best agent isn’t one that can answer every question but one that helps users ask better questions.
If you’re a decision-maker, make “explainability” and “revocability” core metrics. An AI system that can’t explain its decisions and can’t let users roll back operations—no matter how efficient—shouldn’t be deployed in critical scenarios.
QQ Pets twenty years ago taught us: digital companionship can bring real emotional satisfaction, but it can also replace real human connection. Today’s AI agents face the same proposition, only with higher stakes—it affects not just our leisure time but our work output, cognitive habits, and social structure.
In the end, whether technology is good or bad never depends on the technology itself, but on how we use it, and whether we stay clear-headed enough to safeguard the independence of “wisdom” amid the convenience of “intelligence.”
Some technical details in this article are based on public materials and reasonable speculation; corrections are welcome. The experimental code for the personal agent will be open-sourced after it’s organized.
References and Further Reading
- Model Context Protocol specification
- The Coming Wave, Mustafa Suleyman
- My previous article: “The Automation Paradox: Why the Smarter the System, the More It Needs Human Judgment”
评论
0 条评论