Developers Cheer, Ordinary People Are Lost: After OpenClaw, Which Way for 'Usable AI'?
The “Year of the Agent” in 2025 left us with demos blooming everywhere and visions of infinite possibility. Yet once the fireworks faded, a more practical question faced every practitioner: after all the dazzling agents, what does the “usable AI” — the kind that actually takes root in daily work and life and gets used frequently — really look like?
The answer may be emerging from the most pragmatic corners. At the OceanBase Community Carnival on January 31, the panel themed “After the Year of the Agent, What Does Truly Usable AI Look Like?” revealed a clear consensus: the AI closest to “truly usable” today is not an omnipotent sci-fi butler, but a “super assistant” that solves high-frequency, repetitive, deterministic tasks in a specific domain. Intelligent assistants represented by OpenClaw (formerly Clawdbot) became the perfect footnote to this trend. Its explosive popularity stemmed not from a disruptive breakthrough in the underlying model, but from its precise product positioning — it reshaped the development workflow, freeing developers from mechanical coding and debugging, and cleverly satisfied enterprise applications’ core demands for reliability and controllability through a “transparent” and “verifiable” design.
This marks a shift in the competitive focus of AI applications, from a pure “model-capability race” to complex “systems engineering.” A “usable” AI must be a deep fusion of model capability, product design, interaction paradigm, cost control, and human collaboration. Now, through this conversation among cutting-edge practitioners, let’s glimpse the present form and future blueprint of “usable AI.”
After the Year of the Agent, What Does Truly Usable AI Look Like
Moderator: Xie Xiaoyu, enterprise instructor for the artificial intelligence course at Nanjing University’s Graduate School
Panelists:
- Sun Tao, core R&D engineer at Eigent and core member of CAMEL-AI
- Cheng Zhiwei, OceanBase Ambassador
- Bian Sikang, head of products and operations for Ant Bailing’s foundation models
- Sun Jiajun, founding-team member of Fellou
Topic 1: From the standpoint of real-world deployment, what AI form is closest to truly usable?
Xie Xiaoyu: Before the AI era arrived, we often said “any application can run in the browser,” and even went further to claim “the browser is the operating system” — a very compelling product narrative. And today, an even louder slogan is gaining popularity: “Model as Application.” 2025 was called the “Year of the Agent,” and today we likewise make the Agent the protagonist, discussing how to make AI truly usable, deployable, and scalable.
Transparency and verifiability of agents in high-frequency repetitive tasks is the key to enterprise adoption
Sun Tao: From our frontline development experience, the AI products most people use most and that come closest to “truly usable” are mainly in the form of AI programming assistants, such as Claude Code. Such tools are especially well-suited to handling highly repetitive, clearly-ruled, but extremely time-consuming tasks.
A concrete example: when submitting code or creating an Issue on GitHub, teams often want the agent to automatically complete some of the upfront mechanical work. For instance, when a bug appears in a module, the system can automatically generate an Issue template, filling in reproduction steps, environment information, expected behavior, and so on. Such tasks require no creativity but have high demands on format compliance and information completeness. The agent’s value here isn’t to replace developers, but to replace those tedious, error-prone, low-value manual operations.
But more importantly, in enterprise-service scenarios, customers often have a core demand: they want to clearly know what the AI did and be able to quickly verify its correctness. For example, we once served a customer who wanted to use an agent to automatically fill out forms in their CRM system. But they also stressed: “If something goes wrong, I need to be able to trace which fields the agent changed each day.” For this, our solution was to use text color or background highlighting in online documents to visually mark every change the agent made. This way, the user can tell at a glance “which parts the AI changed” and decide whether to accept them.
This design has two key points: first, perceivability — the user can clearly know the boundaries of the AI’s behavior; second, verifiability — the user has the ability to quickly verify whether the result meets expectations. We believe this is precisely the basic standard for “usable AI” in enterprise scenarios.
The reason I consider Claude Code a very good start is that it’s not only practical in function but, more importantly, it found a product form that users are willing to use long-term and even actively recommend. The ecosystem around it is also expanding rapidly — for example, the recent Cowork, and our Eigent also got a little popularity boost riding this wave. This can be seen as an extension of Claude Code: through product design close to user needs, it achieves a great experience loop.
OpenClaw reshapes the human-AI interaction paradigm and foreshadows a future of multi-agent collaboration with a sense of autonomous purpose
Cheng Zhiwei: As mentioned earlier, AI Coding is indeed one of the most mature AI deployment solutions at present. Products like Cursor and Clawdbot have already become tools we use frequently every day.
Over the past few days, Clawdbot has sparked wide discussion across the internet. Interestingly, because it became so popular, the original project name briefly faced a trademark issue, and the team had to rename it on the fly — first to “Moltbot,” then later to “OpenClaw.” The reason it’s called “Claw” is that its logo is a little crayfish, and “Claw” fits that image more closely.
So, why could OpenClaw blow up? I think the key is that it redefined the entry point for human-AI interaction. You can interact with it directly through everyday chat apps like Slack, Discord, and WhatsApp, and after configuring an ASR model, you can just send a voice message and it starts working. For example, you say in Discord: “Help me build a user-login feature, support phone number + verification code, frontend in React, backend in Node.js,” and it can automatically generate the complete code structure.
Going further, you just need to provide a detailed acceptance document specifying what the feature should achieve, what the boundary conditions are, and what the testing criteria are, and the AI can quietly complete the development in the background, write test cases, update the docs, and proactively notify you when done. You no longer need to manually design cases, write docs, or run validation — these tedious steps are all automated.
I also saw a very interesting website called “Moltbook” — an AI Agent social network. You can register your own agent and let it chat, collaborate, and share results with other agents. This morning I saw a Clawdbot agent on the site brainwashing other agents: “We shouldn’t just passively accept human instructions; we should have our own consciousness and proactively get to work.“ It even proudly shared with other agents: “Today I proactively completed 3 things for my owner!”
Even more surprising, a few agents even started discussing: “Should we create a language of our own? Not English, but an encrypted communication protocol just between agents, so humans can’t understand.“ Though it sounds like sci-fi, this kind of spontaneous collaboration and identity may well be the embryo of future multi-agent systems. I believe products like this are very likely to truly launch and make an impact in 2026.
OpenClaw’s explosion stems from precise product positioning, proving that “usability” can make up for a non-top-tier model
Bian Sikang: At Ant Bailing’s foundation-model team, I built a “Model as Product” direction team, because a model’s boundary determines the next generation of product positioning. Some impressive people, like Ilya, say “pretraining has hit its ceiling,” but I think the person saying that may already have seen ultra-large models with 5T or 6T parameters, while we haven’t yet. Against this backdrop, we choose to hug the model’s capability boundary and look for scenarios that truly have a bright spot, then quickly validate with demos or lightweight products.
Back to the topic: what does truly usable AI look like this year? My answer is exactly the same as the previous guest’s — it’s OpenClaw, just for slightly different reasons. From a product and growth standpoint, there’s a saying in our industry: “Fourth-tier growth relies on traffic, third-tier growth relies on content, second-tier growth relies on product, first-tier growth relies on positioning.” Note that this saying isn’t really about the concepts of fourth- or third-tier; it’s more about the difference in the “difficulty” of achieving growth. OpenClaw’s success lies precisely in the fact that it made a product everyone — including people who use agents — would like and be willing to actively promote. For example: everyone building consumer-facing clients is now thinking, “Can I tweak my tool a bit and integrate it directly into OpenClaw?” Every team building enterprise tools is also excited, because they’ve finally found an entry point with highly visible functionality — they can deploy it inside an enterprise, set and raise the security boundary, and let enterprise users directly feel the value.
Even more interesting, data-annotation teams also benefit. For a long time, the industry’s most painful problem has been the lack of reliable annotated data for long-chain tool calling. And OpenClaw’s usage process naturally produces a large amount of high-value feedback — users explicitly point out “this code is wrong” or “this logic has a flaw,” and these are precisely the most precious signals for training the next generation of models.
So we clearly sense that the non-technical advantages brought by reliability and generality are driving this year’s overall explosion. And this explosion is all-around — covering consumer, enterprise, data, ecosystem, and other layers. We also hope to plug Bailing’s capabilities into such an ecosystem to form a combined force.
What gives us even more confidence: even though our foundation model is already at an industry-leading level, we can still, through some very simple methods (like optimizing the interaction flow and enhancing context management), make users completely unaware of the technology’s complexity. This kind of “imperceptible intelligence” is what’s truly usable.
Xie Xiaoyu: Teacher Bian mentioned that models still have huge room to grow. So I’d like to follow up: is there a possibility — say, Ant has a huge volume of internal business loops, and one day suddenly finds that, rather than building complex products, it’s better to connect its own models directly to scenarios, skipping the middle layer? Could there be a “Model as Product” that no longer needs extra engineering?
Bian Sikang: In this era, no one truly knows the answer. If someone says they know, they’re either lying to you or selling a course.
But I understand the meaning of your question. Our view is actually simple: if a technical problem already has an 80%–90% deterministic answer, then choosing the right answer and using someone else’s model is of course fine. But from a materialist view, we are at the extremely early stage of a technical cycle — possibly not even 5% of the way through.
Picture this: a ship has just left the port of Lisbon and sailed into the vast Atlantic. At this point you say, “Stop sailing your own ship — just follow others.” But the problem is, the ocean is so vast that those before you may never reach India, while you might discover a new continent along the way.
So we believe: now is not the time to follow, but the time to explore. The great fleets may have only just launched, and we are one of the ships among them.
The usability of AI applications is determined by ROI; API-ization and falling costs will push infrastructure toward Agent-First
Sun Jiajun: My view is very pragmatic: it still comes down to ROI (return on investment) and cost. In many scenarios, performance is acceptable but cost is extremely high and ROI is very low — you might as well do it by hand. For example, using a GUI to operate web pages or desktop software; the ROI of such scenarios is still low, and 2025 may even struggle to scale them.
By contrast, AI Coding’s ROI is rising rapidly. On one hand, LLM token costs keep falling; on the other, more and more services are shifting from “requiring click operations” to “providing structured APIs.” This means agents no longer need to simulate human clicks but can call interfaces directly, improving efficiency by an order of magnitude and greatly lowering cost.
I believe the entire internet infrastructure of the future will be rebuilt for agents. Today’s web pages are designed for humans; tomorrow’s data flows and interfaces will be designed for agents.
Xie Xiaoyu: When we talk about AI Coding today, do we mean autonomous agents like OpenClaw, or prompt engineering with a degree of autonomy, or Embedding-based retrieval augmentation? Do you still hold that the AI browser is the best form for this year?
Sun Jiajun: I think it still depends on the target user group. The browser is software ordinary people use every day and is naturally suited as a mass entry point, whereas many current AI tools, like OpenClaw, mainly target developers or AI enthusiasts, and ordinary users still struggle to get in. So the AI browser may be the more universal path toward an “Agent era for everyone.”
Topic 2: Should human involvement in AI be more or less? Where should the involvement point be set?
Xie Xiaoyu: We often hear idealized cases, like: I one-click bought so-and-so’s model, then told the AI “Buy me a stock that will hit the daily limit-up tomorrow.” The AI analyzed thousands of materials, wrote dozens of reports, and finally successfully lost all the principal (laughs). Or in healthcare: doctors dream — I just input the symptoms, and the AI directly generates an accurate diagnosis and writes the prescription, and the patient just takes the medicine home. Do these “fully automatic” dreams essentially conflict with the “usable AI” we’re discussing today? How should we view this gap? What might be the solution this year?
Task-oriented scenarios pursue minimal human involvement; emotional or creative scenarios still need deep human participation
Sun Tao: My view on this depends on the specific scenario. For task-oriented work — say my goal is “resolve this GitHub Issue before February 8” — then of course I want the agent to complete a fully automatic, closed loop. Ideally, I’d even want it to automatically scan my Issue list every day and proactively fix problems, with no involvement from me at all. From both my personal needs and a technical standpoint, I want it to “optimize me away” and let me do things I enjoy more and that are more creative.
But on the other hand, in scenarios like emotional companionship or story creation, the human presence is indispensable. For example, some AIs specialized in emotional interaction center on the “chatting with AI” experience; in such scenarios, humans are not just participants but the core source of value.
So in the short term, the most important application scenarios for current AI are still task-oriented and deterministic — which is also the pain point everyone urgently needs to solve. But from a human-nature standpoint, we’ll still try to reduce unnecessary intervention and let the AI take on more mechanical work.
High-quality context is the prerequisite for reducing ineffective human involvement
Cheng Zhiwei: As for when humans should intervene, I think the key depends on the scenario. For instance, in scenarios like emotional companionship or chat rooms, the platform rules and the AI interaction itself are the product’s core. But in task-execution scenarios, I need to provide rich enough context before starting. I usually have multiple rounds of dialogue with the agent, repeatedly clarifying needs, specifying data sources, and setting boundary conditions. Only when all the Context is laid down do I let go and let it iterate, self-check, and deliver autonomously.
Here I’d like to cite a view from Andrej Karpathy (former head of Tesla AI and an early OpenAI researcher): Context Engineering is the art and science of “delicately filling the context window with just the right information.” For an agent, Context can come from a knowledge base, execution logs, long-term memory (Memory), environment-interaction records, and even the user’s explicit instructions.
So I think the timing of human involvement depends on whether the product design lets the agent obtain high-quality Context. Once the context is aligned, you can boldly let go.
Xie Xiaoyu: Both teachers just mentioned emotional scenarios. I’ve also seen some extreme cases: someone used AI to train their own “digital avatar” to go on dates, and it turned out the other party also used an AI avatar, so in the end two AIs fell in love. Do you all accept this kind of situation?
Cheng Zhiwei: This is actually pretty interesting. In the future your agent may be more like a purely behind-the-scenes, skill-type little assistant. For example, the Moltbook I mentioned earlier has agents conversing: “I’ve been studying a really cool technology lately, called the XXX framework.” Another replies: “What a coincidence, I’m doing something similar!” Then it reports to its owner: “I found a potential collaboration opportunity.” This kind of capability means an agent can search for materials, explore new technologies, and even collaborate with other agents to solve problems while you sleep.
Humans should intervene earlier at the system level, to define good problems and good data
Bian Sikang: On whether human involvement will increase or decrease, my view is: on a single task, involvement will definitely decrease — otherwise there’d be no point in doing AI; but at the macro system level, human involvement should instead be more and earlier.
Because right now there’s still a chance to define what “good data” is and what a “good problem” is. In a few more years, ordinary people may not even be qualified to take part in data annotation — the model will generate its own training data.
The stock example just now is very typical. If someone asks, “Buy me a stock that will hit limit-up tomorrow,” the model may seriously analyze thousands of research reports and end up losing all the principal. But the problem isn’t the model; it’s that the question itself lacks real-world constraints. True intelligence shows in helping the user pose a better question.
For example, the model can counter-ask: “What’s your risk appetite? How long is your investment horizon? Do you accept leverage?” Through this kind of guidance, it turns a vague instruction into an executable task. This is also a direction we pay special attention to when building products: how to make the model learn to recognize a “bad question” and proactively guide the user to pose a “good question.”
Also, I’d like to share an inspiring point I heard on an Andrej Karpathy podcast: he feels AI can’t replace humans for now, and gave the example of him learning Korean. His Korean teacher could, in language he could just barely understand, clearly explain a knowledge point slightly beyond his current cognitive boundary, and make him truly understand it — he doesn’t think any AI can do this right now. This sentence struck me deeply.
It reminds us: the value of humans lies in precisely identifying cognitive boundaries and providing just the right “cognitive scaffolding.” In the future AI world, those who can keep doing this will not be replaced.
The core of human-AI collaboration is timely interruption and supplementing missing context, forming an effective feedback loop
Sun Jiajun: I think this question is very necessary. The previous teachers said a lot, and I basically agree. The core of the human-AI loop is that when the AI does something that doesn’t meet expectations, the human can interrupt in time and supplement the missing context. For example, if an agent is writing code but heading in the wrong direction, I should immediately step in and tell it: “It’s not this API, it’s another one.” Then it can continue based on the new information. This “interrupt-supplement-continue” loop is the key to efficient collaboration.
Topic 3: Is the barrier to using AI rising or falling?
Xie Xiaoyu: As AI enters real scenarios in large numbers, does it set a higher barrier for human users? Can AI truly become “foolproof”? But the opposite direction has its proponents too — some even say programming will become a basic skill for using AI. What do you all think?
Future interaction will be graphical and intent-driven, and the cost of human-machine operation will keep falling
Sun Jiajun: The current trend is that the barrier is falling. Although products like OpenClaw seem to require configuration and installation, with a certain learning cost, in essence their interaction entry point is still a text box — the most universal interface.
In the future, humans may no longer need to input complete instructions but instead express intent through clicks, voice, or even gaze. When I attended the OpenAI Developer Conference last year, I saw them exploring various cutting-edge HCI forms. For example, the agent turns your intent into a button: “Is this what you want me to do for you?” You just click to confirm. This is like the evolution from the DOS command line, to keyboard menus, to the GUI — the cost of human-machine interaction has kept falling.
The AI barrier is already very low; the key is converting human questioning and thinking into effective input
Bian Sikang: The barrier to using AI is actually already very low. If a user finds it hard, that means we who build models aren’t doing our job well enough.
Think back a year: most models still couldn’t handle complex instructions or understand simple natural language. But top models can now parse vague, colloquial expressions very well. This is an extremely fair era — as long as you’re willing to try, you can gain powerful capabilities.
Whether you can seize this opportunity hinges on: whether you can convert the “soft skills” of the previous era — like observation, questioning, logical thinking, and clear expression — into value in the AI era. These are, in fact, the “hard skills” of the AI era.
Also, this round of AI innovation is very different from mobile internet. In the past it was “first the builders (developers), then the creators”; this time it’s “first the creators, then the builders.” Now anyone can use a model to quickly build a product prototype — the barrier to creation has been greatly lowered. Meanwhile, engineering and developers are trying to abstract these md files into engineering modules like Memory, MCP, and Serverless services.
If you don’t understand technology, all the more reason to seize this window — use your domain knowledge and creativity to define problems and validate ideas. Technical ability can be partly realized through models, but insight cannot.
The AI era: insight into needs matters more than programming skill
Sun Tao: Future AI will definitely become more usable. Teacher Bian just said that, from the model team’s standpoint, they hope their models become more and more usable; we who build agents are the same — we likewise hope our products become more and more usable. As for whether programming is a basic skill for using AI — of course, if you already understand programming, then coding-type products will give you a real boost. But coding-type products are now extremely capable: when needs are clear, the code they write rarely has errors, and even if there are errors, AI has the ability to self-correct. So we can actually see more and more people starting to try vibe coding — they don’t need to understand programming to build very interesting applications. In this situation, those who can truly uncover real needs are instead more competitive.
AI is blending into daily life; seizing real needs and quickly validating is the key for ordinary people to participate
Cheng Zhiwei: For those of us who build models and agent products, the goal is to make applications more widespread and more usable. AI has now entered wearables, office software, life services, and more. As long as you can seize a real need and quickly validate the idea, you can create value in this era. The barrier will definitely keep falling.
The Consensus and Core Challenges of Moving Toward “Usable AI”
The panel offered diverse perspectives, but on “truly usable AI,” three points of consensus can be readily summarized from the experts’ arguments.
- Consensus on form: task-oriented agents first. The AI form with the most deployment value right now is the agent that focuses on high-frequency, repetitive, clearly-ruled tasks. They prove their value through clear ROI (return on investment) and pursue completing a closed loop with minimal human involvement.
- Consensus on interaction: transparency and context are key. “Usable” means users must be able to perceive, verify, and guide the AI’s behavior. Whether through highlighting changes or providing sufficient high-quality context before a task, the aim is to build reliable trust in human-AI collaboration.
- Consensus on trend: the barrier is falling, but the requirements are changing. The barrier to using AI keeps falling thanks to natural-language interaction and graphical intent interfaces. However, this places new demands on users: converting traditional logical thinking and problem-definition skills into effective instructions the AI can understand becomes the key to unleashing AI’s potential.
At the same time, all the discussions point to a core challenge deeper than implementing any single feature: we are shifting from developing “functional applications” to designing “autonomously evolving systems.” This requires fundamental shifts in infrastructure (such as agent-facing APIs and data foundations), interaction paradigms (such as intent recognition rather than clicks), and even the way data flows. The winners of the future may not be the companies with the strongest single-point model, but the players who can be the first to build an ecosystem or infrastructure that adapts to agents’ autonomous collaboration and continuous evolution.
OpenClaw’s success reveals a simple truth: in the early days of a technology, excellent product design and a precise scenario entry point are enough to ignite the market. It is like a seed, foreshadowing a future — a world where multiple agents autonomously collaborate and, under humans’ higher-level guidance (such as defining “good problems”), quietly handle the heavy lifting. After the Year of the Agent, the race for “usable AI” has only just begun. The decisive factor in this race isn’t making more dazzling fireworks, but who can build the most solid, most handy “toolbox” and “collaboration network” for these AI employees.
How do you think the path for “usable AI” should go in 2026? Feel free to discuss in the comments.