Video: Luo Fuli — OpenClaw, Agent Frameworks: The AI Paradigm Has Already Changed Dramatically. Source: Zhang Xiaojun Podcast Link: https://www.youtube.com/watch?v=V9eI-t3TApE
Don’t treat this interview as just a pitch for OpenClaw. Luo Fuli is talking about something far more concrete: the center of gravity in AI competition is shifting away from raw models toward the combined efficiency of models, agent frameworks, post-training, and organizational design. The model is no longer the sole protagonist—the system is.
Here are a few takeaways I jotted down after listening. The quotes are all from the video; the takeaways are my own.
The Model Is the Engine, the Framework Is the Car
When looking at LLMs in the past, people were used to obsessing over parameters, leaderboards, and pre-training—whoever was bigger and scored higher was crowned the winner. In the agent era, the math doesn’t work that way. A mid-tier model placed in a capable enough framework—paired with memory, tools, task decomposition, multi-model collaboration, and long-horizon execution—can pull off things that previously seemed far out of its league.
Capability is no longer locked solely inside model weights; it’s also embedded in system orchestration. The engine matters, of course, but if the transmission, suspension, navigation, and driver-assist are all garbage, even the most powerful engine will just spin its wheels.
A Complete Reversal in Three Days
The part where Luo Fuli talks about OpenClaw is the most interesting. At first, she was dismissive, writing it off as just Claude Code bolted onto instant messaging, maybe even tainted with marketing hype and hand-waving. A lot of technical people react like this: the second they see productized packaging, their instinct is to dismiss it as impure.
Once she actually got her hands on it, her perspective shifted day by day:
- Day one: she felt it had a “soul.” In plain terms, the context orchestration was meticulously crafted; when small details like time, memory, prompts, and interaction patterns are stacked together, the entire experience transforms.
- Day two: she started delegating non-coding tasks to it, like team management, screening for curiosity, and organizational design.
- Day three: she handed it research tasks and discovered it could co-construct User Agents and fit right into post-training data and RL workflows.
It took a researcher just three days to go from looking down on a tool to overhauling her judgment. That’s more convincing than any product marketing.
The Value of an Agent Framework Is Covering Weaknesses
She said something I strongly agree with:
The greatest value of an agent framework isn’t showing off fancy tricks—it’s covering the model’s weaknesses.
When discussing AI, many people still like to ask, “Which model is the strongest?” It’s not an invalid question, but it’s no longer enough. The question we should be asking is: in a specific task, can the model be reliably amplified by the system? Can it retain long-term context, find its own tools, turn failures into lessons, and transfer conclusions from one task to the next?
If it can’t do that, it’s just a chatbot. If it can, it finally starts looking like real productivity.
Why It’s Easier to Blow Up in the Chinese-Speaking World First
My take is that the hunger for efficiency is much stronger in China. Anything that saves headcount, saves time, or turns a single person into a small team gets pushed into real-world scenarios much faster here.
That’s also where the risk lies. Many people treat agents like some magic silver bullet for automation, only to run the new tools with their same old processes. Agents force you to rewrite your workflows: checkpoints that once required human supervision get inspected autonomously by the system; task delegation that once relied on gut intuition shifts to scheduling driven by memory, context, and feedback loops; managers asking for status updates gets replaced by agents reporting exceptions on their own. If the organization doesn’t adapt and just tacks on an AI tool, the efficiency gains will be minimal.
Post-Training Is Becoming the Main Battleground
Luo Fuli touched on GPU compute allocation. She noted that a reasonable ratio across research, pre-training, and post-training might be 3:1:1.
What lies behind this ratio is that in the agent era, post-training and research exploration are no longer just cosmetic polish applied before a model ships. Pre-training used to be laying the foundation while post-training was interior decoration. Not anymore:
Post-training today is more like converting a house into a self-operating factory.
You have to train it how to use tools, how to plan across long horizons, how to interact with its environment, and how to recover from failures. This directly alters resource allocation—and it reorganizes teams.
Big-Company Resources Don’t Equal Speed
Building agents requires multi-system coordination: models, data, RL, infra, product, toolchains, and evaluation all have to move in sync. If a team is still split across traditional big-tech functional silos, it will quickly grind to a halt. Luo Fuli mentioned that even though MiMo and MiClaw sit inside a tech giant, they effectively operate like a startup.
Big companies have resources, but resources don’t automatically translate to speed. The true rarities are judgment and reaction time.
Proving Yesterday’s Self Wrong Every Day
She had another line:
Every single day, you’re proving yesterday’s self wrong.
Amid this wave of AI shifts, this isn’t cheap self-help fluff. A judgment you formed yesterday might expire today; a tool you just mastered gets absorbed by the framework tomorrow; a capability you thought demanded human touch gets reliably half-automated by an agent the day after. The most dangerous people right now are those who’ve toyed with AI a little and assume they’ve got it all figured out.
Three Things That Will Widen the Gap Going Forward
The gap won’t come from whether you know how to write prompts, nor will it come from subscribing to the most expensive models. It will come down to:
- Whether you can break tasks down into systems rather than one-off Q&As;
- Whether you can crystallize experience into skills, memory, and workflows rather than starting from scratch every time;
- Whether you can accept that your process from yesterday is already outdated—and overhaul it immediately.
The biggest takeaway for me from this interview is that agents aren’t just another product category—they’re a whole new way of organizing work.
It’s the Same Story for Content Production
In the past, content teams competed on individual talent: who picked the best topics, who edited the cleanest, who wrote the sharpest copy, who understood the algorithms. The next phase is about competing on systems: how trending topics are captured, how assets are ingested, how scripts are generated, how audience comments feed back into topic selection, how viral structures are codified into templates, how personas across different accounts are isolated, and how distribution across platforms is handled.
Humans still matter, but human value will concentrate heavily on judgment, taste, trade-offs, and taking responsibility. Repetitive labor will become increasingly worthless.
Stop Just Asking “Which Model Is Stronger”
What really struck me about this interview is that she treats AI neither as magic nor as run-of-the-mill software. She examines how system boundaries are shifting: which capabilities stem from the model, which from the framework, which from organization, and which from environmental feedback.
Does my current way of working need to be rewritten by agents? That’s the question I’ve been turning over in my head ever since listening.