OpenClaw’s explosive popularity has once again exceeded expectations. Alongside the recent debate over Big Model versus Big Harness, the speed of technological development has left me both impressed and somewhat uncertain. After several recent conversations with people in the industry, I have formed some preliminary views on the three elements that matter most to AI today: foundation models, harnesses, and tools. I would like to share them here.
A first principle of AI: harnesses will eventually be dismantled, and large harnesses are not the technological end state
This has always been my view, but the latest OpenClaw wave has challenged it considerably. I began to wonder whether a harness could grow into an ecosystem of its own, forcing foundation models to adapt to it instead. This may no longer be merely possible; to some extent, it is already reality. There are several obvious examples. Commercially, OpenClaw previously clashed with foundation-model companies, demonstrating that the “harness” itself had acquired considerable bargaining power. Technically, foundation models that better support Skills and related memory systems have shone during this wave. We can even expect major vendors to optimize their foundation models specifically for OpenClaw.
Nevertheless, I still hold my earlier judgment and regard these as short-term trends. A harness is fundamentally a patch for a foundation model, and there is no reason its capabilities cannot be absorbed into the model. Foundation models will continue to improve their long-term memory, process control, and tool-use capabilities until they cover most of what today’s harnesses provide.
At the current level of foundation models, harnesses can greatly expand application boundaries and gain a short-term advantage
The main advantage of today’s agents still comes from orchestration and process control that produce results beyond what a foundation model can achieve on its own. In 2023, when foundation models generally lacked reasoning capabilities, workflows for reflection and reasoning formed the core of agent harnesses. Today, their focus has shifted to long-term memory, continual learning, and long-horizon control. Most of the time, harness functionality remains ahead of the model and guides foundation models toward their next generation.
Breakthrough harnesses such as OpenClaw, however, may have a different significance. If a harness develops an enormous distribution advantage in the short term, the capital and data that follow could become critical ingredients for the next generation of foundation models. From this perspective, the current competitive landscape remains highly uncertain.
Make models adapt to tools, rather than making tools adapt to models and agents
OpenClaw’s popularity has also sparked discussion about “AI-native software.” One notable argument I encountered today was that “future software should be designed specifically for AI rather than humans” and that “the GUI is dead.” This view is not entirely without basis. Search is a typical example. Search products were once designed primarily for users: the appearance of the homepage mattered, and logos and search boxes were carefully crafted. Today, AI-generated search traffic has already surpassed human search traffic, weakening many of the traditional advantages of search products.
Yet not every piece of software or every tool must adapt itself to AI. I believe the GUI is unlikely to disappear; for many applications, graphical interfaces retain inherent efficiency advantages. Future AI systems will not be limited to understanding code and Markdown either. Similar assumptions have appeared repeatedly throughout the history of AI. Should autonomous-driving AI adapt to roads designed for people, or should roads be rebuilt around the limitations of current AI through vehicle-road coordination? Should household robots generalize better to unstructured environments, or should every home be remodeled into a more structured environment that suits today’s robots? These questions have fairly clear answers now, but only five years ago I frequently debated people who held the opposite view.
Looking back at the development of large models over the past three or four years, progress has been rapid, but a recurring pattern has become increasingly clear: harnesses built on foundation models expand the boundary; foundation models quickly follow, absorb those capabilities, and then surpass the previous generation of harnesses. Whether this pattern will continue indefinitely, of course, remains difficult to say.