The Action Loop Will Decide the Agent Era

Models will keep improving. But now the real race is who can close the loop from intent to execution, at scale.

Hi all, I was down with a gross cold last week, so I'm late with my take on Claude Code and its implications for agentic AI. Anywho, here it goes…

---

The next stage of AI is not about smarter models. It is about whether agents can actually do things and quietly take some of the mental load off our lives.

Over the last few weeks, a viral essay in The Cut argued we should embrace more friction and reject AI. Slow down, do things manually, reclaim the human experience. But it wasn’t just that, it was like going to the farmers' market and cooking a gourmet meal from scratch. And honestly, I truly would have applauded this a few years ago, but I was so triggered reading this.

There is something comforting about the idea that the “right” life is the one with more texture, more effort, more intention, fewer shortcuts. But the “friction-maxxing” thesis only works if friction is optional. And only when you have the luxury to farm-to-table, given you have money, have time, and have the proximity to such establishments.

For most people I know, especially anyone raising small kids, managing a household, caring for parents, on top of working a full-time demanding career or juggling too many responsibilities, friction is not some artisanal choice. It is the background noise of life. Inventory, schedules, logistics, coordination, remembering. The endless “small” things that are not small when they add up to your mental load and eat up your capacity, and because they never stop. So friction-maxxing seems like an euphemism for privilege.

A Googled image of Super Mom

And the part people do not talk about enough is that it is not just the physical work. It is the invisible admin layer. The mental overhead you carry around all day, like tabs left open in your brain. You did not choose to spend your Sunday thinking about diaper sizes and snack supplies, medicine refills, calendar conflicts, bills piling up, and the dog food running low. It just happens. And if you drop one ball, then chaos spirals into something bigger and bigger. But that is actually the experience for many adults, if not, most adults.

In no way am I anti-humanity, but with the backdrop of agents taking the world by awe and a realization that the tedious, repetitive, and not-so-fun things can be taken care of in the background, I felt like the preach was a bit tone deaf. If anything, if AI agents can help us with mundane repetitive tasks, this is not outsourcing the intellectual interactions, not outsourcing the ability to smell the roses and cook with your kids. It’s actually the ability to free up so much of the time that weighs you down from being able to do the above- living.

And if you look at how the ultra-rich live, this is the first thing they outsource and get their time back: assistants, drivers, nannies, house managers. For everyone else, there is now a potential “assistant” in the form of AI.

This is why I think we need to shift the way we frame AI. The question is not how intelligent the model is. That brings me to this: what is the action loop now? What can it reliably do and complete for me?

---

What the action loop actually looks like

When people say they want “AI help,” they often mean suggestions. A better search engine. A better autopilot for writing. A chatbot that feels clever. And fine, that is what the whole article was saying we should reject, “don’t get AI to suggest what to eat, what to do,” but what if it’s not really that? What if AI can help really manage certain action loops like weekly orders of milk and sourdough?

Anyway, now that I’ve seen what Claude Code can do, I want something that is painfully practical: an agent that can notify me, ping me, set order for, and help me with some of that mental load.

That is an action loop: context, decision, execution, confirmation.

Systems, not just intelligence

This is why Omar Khattab’s point on the a16z podcast really landed with me a few weeks ago. His framing, paraphrased but faithful, is that it is safe to assume models will keep getting better. The real question becomes what problem you are actually trying to solve.

We do not build systems because we lack intelligence. We build systems because we need reliability.

A powerful model that can do something is not the same as a system that does it consistently, safely, and with the right guardrails. The AI conversation has been stuck on capabilities, benchmarks, scores, impressive demos, when most real-life utility depends on something far less glamorous: permissions, integration, predictable execution, and an interface that makes humans trust it.

That is what moves AI from “wow” to “useful.”

We’ve written about ecosystems, systems, and them all. But seeing AI as a system is a pivotal lens. Rui Ma recently tweeted about Davos conversations circling abstract questions like whether AI has “human-like consciousness” or whether it will replace humans, while Microsoft CEO Satya Nadella was talking about factories, energy efficiency, total cost of ownership, and “how many tokens can be produced per watt of electricity.” The claim is essentially that future AI competition looks a lot like a battle over the efficiency of token factories (than whatever the battle is we’re arguing about right now).

Rui Ma@ruima

Just read a Chinese tech media piece on Davos and AI that I largely agree with. I won’t repost the whole thing, but the core argument is very straightforward. While many Davos conversations were still circling abstract questions like whether AI has “human-like consciousness” or

6:30 AM · Jan 26, 2026 · 4.33K Views

---

4 Replies · 15 Reposts · 96 Likes

If you remove the noise and rhetoric, it becomes an industrial story and a system planning story. If you think about it, it’s about building infrastructure, driving unit costs down, comparing efficiency, setting up guardrails, and then integrating into the real economy. If one company can produce tokens at one-tenth the cost of another, it will probably ‘win’. That is a useful mental reset. End of the day, most real adoption of technology is often constrained by something far more boring: cost, reliability, and the ability to run at scale.

And yet, there is a twist. Even if token production becomes dramatically cheaper, tokens may stop being the thing that differentiates. Once everyone can make tokens cheaply, the question shifts again. Not “who can generate the most,” but “what can they actually get done, how fast, and how reliably.”

It’s an industrial system built that will change the way we operate.

The accessibility inflection

Just like all you wild cats, over the last two weeks, I have been experimenting with Claude Code, Anthropic’s coding agent that lets you build software through natural language. If you have not tried it, the idea is simple. You describe what you want in PLAIN ENGLISH, and it writes code and runs it. It showed me our interaction with AI has completely transformed, once again. I think it’s truly a turning point because before this, a person without coding abilities actually couldn’t make anything that useful, just "vibecoding”.

Minions from ‘Despicable Me’

But tbh, I am still learning how to use it well. In the beginning, it feels quite unintuitive, but then once you get the gist of it, it’s like having a dozen minions working for you, especially when they deploy parallel agents. And then the work is teaching them what “good” looks like, and you have to be explicit about boundaries. So much of it is just iterations. My Saturday date night was spent with my husband finetuning each of our agents side by side with some chamomile tea. *true romance

Even early experiments were revealing as to what I could do. One small example: I asked it to analyze my past writing and pull out recurring themes. That is something I can do manually, but never do, because it is tedious. Claude did not just summarize. It gave me structure, identified themes, and connected dots that frankly is done much better than I could have myself. It felt like it cleaned a (very) messy room (brain clutter) and laid everything out on the table.

That gap between “possible” and “actually done” is the real story.

Jasmine Sun, in her Claude psychosis piece, described this feeling well: you do not fully understand coding agents until you build something. Engineers were “AGI-pilled” early because they felt the compounding effect first.

For developers, none of this is shocking (confirmed by friends). For the rest of us, the non-technical majority, the layman, this is the point where the world changes. James Wang puts it adequately in his newsletter, Weighty Thoughts: a technology only experts can use is basically still a prototype. The economic effects start when regular people can pick it up and get value.

He calls this the “boring phase” of AI. Accessible agents doing things for people. Not flashy, not poetic. Just quietly transformational.

And the early signals are already there, as Wang notes via a J.P. Morgan survey: a meaningful share of businesses are paying for AI tools now, and productivity data is starting to show the fingerprint of technology-driven output growth. We are still early, but the direction is clear.

---

What agents actually enable

Here is the simplest way I can explain why agents are different. They collapse the distance between intent and execution.

Instead of opening a tab, searching, copying, pasting, logging in, clicking through menus, confirming, and then updating another system, you say what you want, the agent runs the sequence, shows you what happened, and asks for confirmation when needed. And for the first time, it is starting to feel reachable, even for people who do not code.

You can feel the competition shifting from “smartest model” to “best task completion at a reasonable cost” now in the mainstream conversations, too. That is why the agent product wave matters. When I first tried out Manus, I wasn’t shocked the same way I was when I tinkered with Claude Code, but then I’m a complete newb, or maybe I wasn’t using it correctly. Probably because I was still stuck in the chatbot/research interaction, thinking of AI.

The newly debuted public company MiniMax’s new agent push captured the vibe well in one phrase: 从刷分到会干活, from chasing benchmark scores to actually getting work done.

Even if you have not tried every product, you can see the market moving. Desktop agents breaking out of the browser. Local files and web automation in one place. “Expert agents” that are closer to workflows than chat.

As James put it, “That is the boring phase”. And boring is maybe exactly what we need right now as we continue to push for practical usage, lower token cost, efficiency and real help to humans in the world.

---

The China angle: who already owns the action loop?

Here is what I think a lot of people may still underestimate about the agent era. Chinese super-apps already own the action loop.

Meta has largely been seen as a potential big AI winner as it owns attention. In the agentic era, it has shelled out some serious dollars to buy out Manus as well. Not many companies can compare with its world-class distribution. But in many cases, the “doing” still happens somewhere else. You might discover something on Instagram, but you complete the purchase in a different system. The rails are not native.

We’ve written a lot about WeChat and Alipay, so you know that the platforms are fundamentally different. If you have intent, you can book, buy, pay, coordinate, and transact without leaving the ecosystem. The platform already has the identity layer, payment layer, service layer, and social graph. That is what makes “agentic” real, because the agent is not stuck giving suggestions. It can execute.

This is why Alibaba’s Qwen app launch is interesting beyond the model itself. The thinking is integration: tasks across commerce, payments, travel, maps, and even phone calls for bookings; these are tangible action loops. In the CNBC coverage, Counterpoint’s Shaochen Wang essentially says the quiet part out loud. Alibaba is well-positioned for agentic commerce because it has both the model and the network of services.

That is what “from question to action” looks like when it is not just a demo.

Tencent’s move: context at scale

Tencent’s setup may be even more powerful. Not because the model is magically better, but because the context is. WeChat is not just a utility app. It is a living social graph with accumulated conversation history and group dynamics. It is where family logistics and work logistics already happen. I have always thought that they would have an advantage in the ecosystem, but I overlooked the fact that they have such a strong context base, and it serves so many purposes, from managing households to personal diaries to workstreams.

So with Yuanbao’s most recent "leak,” articles have described how agents could be pulled into everyday coordination. Summaries, reminders, routing tasks to the right person, lightweight admin that currently drains attention inside group chats. [FYI, it was announced today that Pony Ma has greenlit a project over CNY to give up 1 billion RMB worth of red packet money through Yuanbao, the highest amount you could potentially ‘win’ in a single shake can be upwards of tens of thousands of RMB, over CNY to incentivize people into the app]

Tencent’s Yao Shunyu recently made a point that is hard to argue with. A lot of AI lacks economic value today, not because the models are dumb, but because they lack context. They do not know you or your life. WeChat can provide that at scale. And this echoes what he was saying at the AGI Next Summit, or I guess hinting at this point, about what if we can pull in an AI agent to help me with booking a restaurant for my wife and me as we are having a conversation about what we want to eat tonight.

The structural divide is getting clearer

This is the AI super-app thesis we have been circling at AI Proem for a while. The next evolution is not a separate AI app. It is agents embedded into existing touchpoints, where people already live and already transact.

Forrester’s Charlie Dai told CNBC that Chinese platforms benefit from integrated ecosystems, rich behavioral data, and consumer familiarity with super-apps. Bernstein’s “AI Total War” report makes a similar point in analyst-speak (I bet they read my work;p). WeChat’s moat is not just a product. It is coordination at scale, and AI companies are trying to become WeChat. Much, much, much like what I wrote about here.

• WeChat and Alipay have a closed loop. Identity, payments, services, social. Agents can execute.

• Google and Microsoft have a partially closed loop. Workspace and identity. Agents can execute within workflows.

• Meta has attention and social graph, but the final “activity” execution is often external. Insta shop is like a meh business. Agents risk staying suggestion engines. This does not mean Meta cannot build rails. But it does mean the starting position is different. In an agent era, distribution is not enough. The platform needs a place for actions to land.

Agents delivering value?

If 2025 was the year of “agents are coming,” 2026 is truly feeling like the year we find out if they actually deliver actionable results.

The competition is shifting from model intelligence to task completion rate, and more importantly, whether those tasks reduce real friction for real people. Bernstein’s line on “engagement growth” matters, but I would add a more human metric: will agents meaningfully reduce the mental burden of coordination?

Because the boring phase of AI is not about dazzling outputs. It is about whether systems can be trusted to handle the mundane repetitive admin that quietly drains our attention. Household logistics. Scheduling. Follow-ups. The repetitive, tedious tasks.

My personal test remains embarrassingly simple. When an agent can reliably track household needs, sync calendars, coordinate schedules, and execute the small tasks, without me becoming the human API between apps, that is when AI becomes real.

That is why I think 2026 is the year agents have to prove they can take action, not just talk. Agents that close loops, share mental load, and move work forward inside the environments we already use.

That is the boring phase, but it might be the most important phase yet. Can it deliver pragmatic results? Can it drive real economic impact?

Well, James Wang says, we’re really just getting started! What does he know? He’s only just been investing in this space for about a decade and wrote a best-selling book about AI diffusion in the real economy and how we can adapt to the changes AI will drive. Jokes aside, this is worth a read.

AI Gains Starting to Show in the Real Economy

---