There's a pattern I keep seeing and nobody's really talking about it.

I've been experimenting heavily with Claude Code these past few weeks, and more recently OpenClaw (that viral personal AI agent). I use AI a lot. And something has shifted in the past few months that's hard to articulate but easy to feel if you’ve also been using AI actively.

The models haven't gotten dramatically smarter. Benchmark improvements have been largely incremental. But the doing has changed. I’m no longer just chatting with these things. They're acting on my computer, writing code across files, running tasks, making decisions in sequences that would've required me to sit there clicking through each step.

Now that’s all and well for personal productivity. But it got me thinking about a bigger question. How does this translate into real economic value out in the world?

Everyone's focused on which model is smartest. Who's winning benchmarks. Whether it is Claude, GPT, or Gemini pulling ahead. And look, that stuff matters. But the more I watch where agents are actually doing things in the real world, not just talking, the more I think.. the model isn't the most interesting variable.

The interesting variable is whether the agent sits inside a closed loop.

By closed loop, I mean something specific. Identity, permissions, state, and execution all living inside the same system. When those pieces come together, agents become true operators. When they don't, you end up with a very eloquent assistant who still needs you to go click the buttons.

CNBC ran a piece a couple of weeks ago framing the entire Chinese super-app competition as an "agentic commerce" race. Alibaba, Tencent, ByteDance, all racing to turn chatbots into full-service transaction tools. And the examples are.. kind of proving the thesis in real time.

Start with Alibaba. On January 15, they upgraded their Qwen chatbot so users can now complete entire transactions inside the chat. Order food through Taobao Instant Commerce, book travel through Fliggy, pay through Alipay. All without leaving the conversation. Promotions get applied automatically, payments settle in-chat.

Alibaba VP Wu Jia put it pretty directly:

That framing is exactly right. And this magical consumer experience is only possible because Alibaba already owns the full stack. Identity, commerce, payments, logistics, maps. The agent rides on top of their proprietary infrastructure that was built years ago for other reasons.

Tencent's been on a similar path. Martin Lau (Tencent’s President) said on their Q1 2025 earnings call that WeChat's ecosystem gives them the opportunity to build agentic AI that's fundamentally different from standalone chatbots because it connects transactional and operational capabilities across many verticals. They integrated their Yuanbao chatbot directly into WeChat, and by mid-2025, it was the fastest-growing AI assistant in China.

In January, WeChat launched a year-long growth plan for AI-powered Mini Programs, providing cloud resources, computing power, and traffic incentives through all of 2026. They're clearly betting that the agent layer lives inside their ecosystem.

(I realize I'm spending a lot of time on Chinese super-apps here. Bear with me. There's a reason.)

Lau has said publicly that the ideal endgame is for WeChat to launch a full AI agent that understands your needs through the social ecosystem, accesses services through Mini Programs, and completes payments immediately.

My takeaway: for consumers, AI agents are the lubricant that bridges the gap between what you want and what actually happens.

Here's the example that really crystallized this for me. I actually covered this story in a previous edition of our newsletter, but a quick recap:

In December, ByteDance launched what many were calling the first truly agentic AI smartphone. ByteDance's Doubao assistant embedded directly at the OS level. Users could activate it with a side button and watch it autonomously navigate apps, book restaurants, compare prices across platforms, edit photos, make purchases. All through voice commands without touching the screen.

The phone sold out instantly. About 30,000 units, resale prices jumped 40% above retail. Social media was full of videos showing the thing completing complex multi-step tasks across different apps. It looked like the future. If I were living in China (i’m not), I would have jumped to buy this.

Then reality hit.

Within days, WeChat suspended accounts that tried using Doubao's voice control. Alipay restricted the AI's access to financial services. Taobao blocked it. Gaming companies complained about unfair advantages. Tencent CEO Pony Ma reportedly called the Doubao phone "extremely unsafe and irresponsible" at a company meeting. ByteDance had to roll back features fast, disabling interactions with banking apps, preventing the assistant from claiming promotional rewards meant for humans, suspending AI features in competitive games.

This is the closed loop thesis playing out in real time.

ByteDance tried to build an agent that operates across closed loops from the outside. And the platforms that own those loops shut it down. Not because the AI wasn't capable. It clearly was. But because it didn't have native authority inside those systems. It was trying to act without permission, without identity, without being a trusted participant in the loop.

ByteDance CEO Liang Rubo said at an internal meeting in early 2026 that "AI is hitting a new peak" and the AI assistant ecosystem is their most critical short-term objective.

Just this week, they announced that they’re preparing to relaunch a second-gen phone for Q2 2026, so maybe I’ll finally get a chance to try it. They’re also negotiating access deals with ride-hailing, food delivery, and ticketing platforms. But the fact that they have to negotiate at all tells you everything about who holds the power. The platforms with the closed loops.

And so I think consumer agents scale faster in China than in the US. Not because the models are better, but because three structural conditions already exist.

First, unified surfaces. Messaging, payments, services, and identity all sit inside a few dominant apps. An agent can act without leaving the environment.

Second, people already trust software to execute. Chinese users are very familiar with using apps to pay, order, schedule, and coordinate on their behalf. Every day. For years. Letting an AI do the same thing is a smaller behavioral leap than it would be in the West, where those actions are scattered across dozens of apps and services.

Third, settlement is native. Payment transactions complete inside the platform. No redirects to a third-party checkout. No friction between wanting something and it being done.

This creates a tight loop. Intent, execution, payment, feedback. All in one place. That loop can scale quickly.

But this doesn't necessarily mean the US falls behind in the economic impact of consumer AI. It just shows up differently.

In the US, the closed-loop story is playing out in two places. Enterprise productivity, and now, commerce.

Google and Microsoft control a lot of the enterprise pieces. Identity (Google accounts), work context (Docs, Gmail, Sheets, Calendar, Office, Outlook, Teams), and workflow boundaries. Inside that workspace, agents can already act. They can schedule meetings, draft documents, update spreadsheets. In many enterprises, the workspace is the business.

Both are integrating with open protocols like Agent2Agent and MCP to let agents collaborate across organizational boundaries.

What they still don't have is a native consumer money rail. Payments, logistics, and external consumer actions still spill out into third-party systems. Their agents feel strong at work, weaker outside it.

So to me, the more interesting recent move is on the consumer side. OpenAI launched Instant Checkout in ChatGPT, where users can now buy directly from Etsy sellers in chat, and soon over a million Shopify merchants. Instacart became the first app to offer full checkout directly within ChatGPT. Users go from meal planning to a delivered grocery cart in a single conversation.

All of this runs on the Agentic Commerce Protocol, an open standard built by Stripe and OpenAI. It's their attempt to close the loop on the consumer side. Not by owning the whole ecosystem like WeChat does, but by building standardized rails that let agents transact across merchants and payment systems.

It's a fundamentally different approach. China closes the loop by consolidating everything on a single platform. The US is trying to close it by connecting many platforms through shared infrastructure.

One country built a mall. The other is building roads between every shop in town.

Different paths, but both lead to agents that can actually do stuff, not just summarize essays like this and give you ‘key takeaways.’

Okay wait, let me be more precise about this. Model quality isn't irrelevant. But it's not the bottleneck people think it is.

An average model inside a closed loop can do much more for the typical retail user than a stronger model that sits outside the system. Because execution beats eloquence. When identity, permission, memory, and action live together, the agent doesn't need to convince you every step of the way. It already has the right to act within boundaries.

And there's a flywheel here that I think is underappreciated. When agents can act, they generate outcome data. Outcomes produce feedback. Feedback updates behavior. Behavior improves execution.

Closed-loop platforms like WeChat will have compounding data advantages in building great agentic consumer experiences. This is why super-apps feel inevitable for agents. They're not just action surfaces. They're learning environments.

Everything I've described assumes the loop belongs to the platform. But there's a scenario where that breaks.

Agents that carry their own authority.

This is already being built. Besides Agent2Agent and Agentic Commerce Protocol, in 2025 we also saw:

If this stack matures.. portable identity, personal data vaults, agent-held wallets, user-controlled permission layers.. the loop moves from the platform to the agent itself. The agent becomes the operating layer across systems.

And then the question isn't who owns the closed loop anymore. It's who controls the agent's authority stack.

I think this is probably the right long-term frame. It's closer to how the internet has historically evolved. Not one closed loop, but many semi-open systems connected through shared rails. The advantage shifts from owning the loop to orchestrating across loops.

But we're not there yet. It’ll take a while for the industry to converge around the standards that everyone actually adopts. It will be a messy process. And so the platforms that already have closed loops have a massive head start on the transition.

If the closed-loop thesis holds (and I'm fairly convinced it does for 2026), a few things follow.

The most powerful AI companies might not be the ones with the best models. They might be the ones that already own closed loops. Or the ones building the authority layer that makes agents portable across all of them.

I don't have a clean conclusion here. The consumer agent layer is going to look very different in China versus the US, and I think both paths are valid but for different reasons. The portable authority thing could change everything, or it could stay stuck in standards-committee limbo for years.

Still working through some of this. But I'm pretty sure this frame is right. The next phase of AI won’t be about intelligence alone. It's about where intelligence is allowed to touch reality without friction.

Cheers,

Teng Yan

P.S. Know a builder or investor who’s too busy to track the agent space but too smart to miss the trends? Forward this to them. You’re helping us build the smartest Agentic community on the web.

I also write a newsletter on decentralized AI and robotics at Chainofthought.xyz.