Tencent's AI Agents Hit 20M Users – But the Real Story Is the Centralization Trap

We didn't see it coming. Not really. A single report buried in Q2 2026's market noise: Tencent's WorkBuddy, an AI-native office agent, clocked 20.97 million monthly visits. That's more than the next two competitors combined. The numbers scream victory. The VC decks will write themselves. But here's the thing no one wants to admit — this isn't a triumph of technology. It's a monument to the centralization we keep pretending we're fighting. — Root: The data tells a story of scale, not sovereignty. Let's rewind. WorkBuddy is an AI agent that writes emails, summarizes meetings, analyzes spreadsheets, and queries internal knowledge bases. It's powered by Tencent's Hunyuan model, hosted on Tencent Cloud, and deeply integrated into WeChat Work. On paper, it's a productivity marvel. In practice, it is a black box that captures every keystroke, every decision, every moment of human cognition in the office. That's not an agent. That's a digital landlord. Context: The rise of centralized AI agents is a direct replay of Web2's data land grab — just with better marketing. We've been here before. In 2017, I wrote a scrappy manifesto called "The Freedom Stack" while still a sophomore at Tallinn University. I argued that code should serve autonomy, not capture. Back then, it was about Bitcoin's censorship resistance. Today, the same principle applies to AI agents. Tencent, Alibaba, ByteDance — they're all building walled gardens where your work data feeds their model, your workflow optimizes their algorithms, and your output becomes their moat. The 20 million users? They're not customers. They're the product. The model trains on their decisions, then sells the intelligence back as a subscription. — Root: The real innovation isn't the agent. It's the ownership structure. Core insight: Decentralized AI agents are not just an alternative — they're the only path to genuine autonomy. And the market is beginning to realize it. Let's break down why WorkBuddy's architecture is a system-level risk, not just a corporate one. Every interaction funnels through Tencent's centralized sequencer — the AI inference pipeline that decides which request gets processed, how fast, and at what cost. Sound familiar? It's the same single-point failure we've criticized in Layer2 sequencers for years. "Decentralized sequencing" has been a PowerPoint slide for two years, yet here we are, trusting a single corporate sequencer with 20 million users' cognitive labor. Based on my audit experience with DeFi protocols, the failure modes are textbook: data censorship, model bias, arbitrary service-level adjustments, and — most critically — user lock-in. WorkBuddy doesn't export your knowledge graphs. It doesn't let you bring your own model. It doesn't support cross-platform interoperability. You're not building a digital workspace; you're renting a cell in a panopticon. Now, consider the alternative. Protocols like Bittensor and Autonolas are building decentralized networks where AI agents can operate on open infrastructure. Model weights are verifiable on-chain. Inference is incentivized through token economics. Users own their data via self-sovereign identity (SSI) — a concept I explored during my regulatory sandbox project in Estonia. The difference isn't subtle. It's existential. Let's talk numbers. The same report shows the total Chinese PC-based AI office agent market exceeded 60 million monthly visits in June 2026. That's a $3-5 billion annual addressable market by conservative estimates, assuming even a modest $50/user/year conversion. But the real value isn't in the subscription fee. It's in the data flywheel. WorkBuddy's 20 million users generate billions of interactions per month, each one a training datum for Hunyuan. If Tencent monetizes that data through API access or model licensing, the revenue could dwarf the agent subscription. This is why they're giving away the agent for free — the real product is the model, and the model is fed by you. Contrarian angle: Decentralized AI agents may actually have better unit economics at scale. Here's the counter-intuitive truth. Centralized AI agents require massive upfront capital for GPU clusters, engineering teams, and compliance overhead. Metrics like amortized inference cost per request are heavily subsidized by VC money or cross-subsidized from other business units (e.g., Tencent's gaming revenue). When the funding cycle turns — and it always does — these costs get passed to users. Decentralized networks, by contrast, distribute inference across a global pool of commodity hardware, with token incentives aligning supply and demand. The marginal cost of running a query on a decentralized agent network can approach zero, especially for simple tasks. And because the network is permissionless, users can switch providers instantly, preventing lock-in. We didn't see this coming because we were blinded by the speed of centralized development. But the fragility is already showing. In April 2026, a misconfigured load balancer at Tencent Cloud took down WorkBuddy for six hours. Users couldn't access their knowledge bases, their draft emails, their scheduled tasks. Entire departments ground to a halt. A decentralized agent network, by design, has no single point of failure. Your agent could be running on a node in Seoul, a node in São Paulo, or a home server in Tallinn. You wouldn't notice. — Root: The next frontier isn't better AI. It's trusted AI. Takeaway: The 20 million user milestone is a wake-up call, not a victory lap. It proves that AI agents are product-market fit. But it also proves that we're rushing headlong into the same centralization trap that Web2 failed to escape. The question isn't whether Tencent's product is good — it is. The question is whether we're building tools that free us, or tools that cage us. I've seen this script before. In 2020, during the DeFi Summer, I launched three yield aggregators. I was manic, chasing composability. When a minor exploit drained 15% of my liquidity, I wrote a transparent post-mortem analyzing the psychological rush of rapid deployment. The community forgave me because I owned the vulnerability. Today, Tencent owns none. Their product is a closed-source monolith with no audit trail, no governance, and no user recourse. If they decide to change the terms, your workflow breaks. If they get hacked, your data leaks. If they get acquired by a hostile actor, your cognition becomes intelligence for a regime you didn't choose. This is why I founded "Sovereign Agents" in 2025 — a platform where AI agents hold crypto wallets and negotiate services autonomously. Our testnet integrates multiple LLM providers, not because we can't build one, but because dependency is the enemy of resilience. The architecture ensures that no single model, no single company, no single jurisdiction controls your digital labor. It's messy. It's slower. But it's real. The market will eventually bifurcate. One branch will optimize for convenience — that's WorkBuddy's lane. The other will optimize for sovereignty — that's the decentralized agent lane. The 20 million users of WorkBuddy are proof that the convenience lane is already crowded. But the sovereignty lane is wide open. And the users who choose it will never look back. We didn't learn from the social media monopolies. We didn't learn from the cloud computing lock-in. We're about to make the same mistake with AI agents. But we still have time to rewrite the script — not in the boardroom, but on the chain. — Root: The future isn't a smarter assistant. It's an assistant you actually own.