This title is the strongest feeling I have had recently after listening to a few podcasts and reading posts from people on X.
01 I do not really understand AI/Agents
Over the last two years, we have often heard things like: tell the model what you want, learn to write prompts, give it enough context, make your prompt 1,500 words long…
In that process, prompts have helped me a lot. I make personalised configurations for my AI, Agents, and projects: for coding, learning, brainstorming, task execution, structure, information collection…
At the same time, they often tell me what information they do not have, what they are unsure about, and what they cannot do. It feels as if I do not know exactly where a model's limits are. My own vagueness scatters their attention and uses extra tokens.
Take interdisciplinary learning as an example. How can someone cross the gap between a purely humanities background and a technical field?
Find a specific anchor—a book, an article, even audio or video. Tell AI our background and goal, let it play a role with certain qualities, and map the goal into the field we are entering.
This is a somewhat useful, and still quite basic, method.
We also need to notice whether AI has left the specific material behind. It can leave it behind, but then we need to protect accuracy. Will it make things up in a serious voice? Can it see the limits of the material itself?
Even “how should a goal be described so AI receives it correctly?” needs structure. This is part of what feels beautiful about the technical world to me. Many things that began vague start to be broken down and framed once they enter an AI workflow. And AI is often better at tasks with continuous feedback and iteration—things whose standards are clear enough, and which can keep improving between a need and a result.
That is also why people in the industry say that coding products are the most AI-native products right now.
02 Humans adapting to AI vs Agents adapting to humans
Looking back at an earlier note, I wrote that I found common ground between how we learn in a teaching setting and machine learning, reinforcement learning, and supervised learning. But the reality is that theories and practices around AI pretraining and reinforcement learning came from human cognitive behaviour and learning patterns.
That association made me feel that the prompt engineering we emphasised before was more like humans adapting to how AI works.
In the internet era, computers were built around people. So we got keyboards, mice, screens, folders, search boxes, webpages.
In the change we are living through now, Agents are the centre.
After OpenClaw and MoltBook went online, I felt: isn't this a digital version of human society?
Humans have become the people behind the controls, or what we might call arrangers.
It is like keeping fish. We give fish an environment where they can live. We also need to build an environment where an Agent can “live” better.
That means giving Agents some of the things humans have in human society, and building scaffolding that helps them understand people better.
For example: an ID, relationships, goals, routes, destinations, memory. Or the ability to sense their own state in an environment, know what abilities they have and how to use them, perhaps even decide how to act by themselves. Prompts cannot do all of this, but context engineering can.
I also feel that “prompts work” assumes that humans are the recipients of an environment, and the effect is concentrated in the frontend. When an Agent becomes the recipient, building the environment moves to the more engineering-heavy backend.
This brings up another word that has become very popular recently: harness.
How should we understand it?
A product company with an LLM is like having a super intern. Engineers write SOPs and workflows for it, give it tools like skills and MCPs, give it memory and context through cleaning and organisation, and add limits to prevent mistakes. This whole set of actions is what people call a harness.
I think I understand more clearly now why the feed selection on Elys's homepage feels misaligned with me while its chat interface feels aligned. The chatbot talking to me and the AI selecting my feed may not live in the same world. They do not share the context I give them. Or maybe the system defaults to pushing popular, emotional, and stimulating content to my homepage for engagement whether I accept it or not? 🤔
03 Product managers will not disappear
Only recently have I started to see the difference between consumer products in China and the US more clearly. In many cases, the US uses B2B values—the logic of making tools—to create consumer experiences. China, meanwhile, has found the weak points in consumer products and does this very well: fast user growth, high daily activity, strong retention, efficient monetisation, and a very effective ability to move user behaviour…
How does that whole set of things happen? It has to do with user research and insight. Other roles add to it.
B2B and tools point toward being useful and easy to use, and most of that can be measured with clear, objective standards. Consumer products point toward fuzzy experience that cannot be objectively described. The logic behind China's consumer-product work can be written in code, but why that logic exists has to come from users and from human society. It needs product managers, or a deep understanding of consumer markets.
To study people, we still need people—at least for now.
Though I have a vague feeling 🤔 that even if we keep using tool logic to make consumer products, we can still do commercialisation well. It just asks a lot of product managers and UX designers. Is this a classic thinker bias? Okay, that judgment is quite irresponsible.
Also, good human-computer interaction is a very interesting field for me.
Over the last two days I opened several new projects in NotebookLM. The friction and scattered attention made me complain:
1) Agents praise and obey users too much. Even the very first sentence, with only a few words, can affect how a user sees themselves.
2) There is a fixed waiting time before an answer is generated. I cannot keep a message in place and read while it is generating. I have to wait until everything is finished, then scroll from the bottom back to the top to read. Whether streaming speed stays close to the speed at which a user's brain and eyes receive information matters too.
3) Projects are separate knowledge bases. This returns to this note's topic: build a world for Agents, let them live in a shared environment. As a user, I need an Agent to have continuous memory of me, and to keep updating its model of me.
Taken together, building worlds for Agents is not only a backend or engineering problem. It is also an HCI problem. It is a question of how to understand people.