Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of Cloud and Smart Industries Group
Spring 2026 marked a subtle but significant shift in the AI industry: the conversation moved beyond just models.
For the past three years, the entire field has been obsessed with model parameters. 175 billion, a trillion, ten trillion—the numbers kept growing, the narratives more ambitious. Each parameter leap sparked another round of discussions about "emergent capabilities." Models learned to reason, to code, to create, seemingly just one step away from true intelligence.
Then, a small lobster stirred the entire industry.
OpenClaw didn't release a new model, didn't break any benchmark records, and didn't train a single new parameter. It did one simple thing: it built a complete working environment for large models—a file system, code sandbox, toolchain, feedback loop, and automated validation. The same model, within this environment, transformed from a mere conversational chatbot into an agent capable of sustained work and autonomous problem-solving.
The only variable was the shell. And that shell now has a formal name: Harness.
Harness, literally translated, means tack—the collective term for reins, bridles, saddles, and harnesses. A wild horse possesses immense power, but without reins, you can't ride it; without a harness, it can't pull a cart; without a saddle, you can't sit securely. The horse's power is raw and potent, and the Harness is the system that converts that force into forward momentum for civilization.
AI Implementation Is an Engineering Challenge, Not Just an Algorithmic One
To understand Harness, we need a clearer framework:
Large models are the engine, Harness is the wiring harness, and the user is the driver.
The engine provides raw power, but the engine itself doesn't drive the car. Similarly, a large model can think and converse, but it lacks operational capability. It needs a wiring harness—a system that transmits power to the wheels, signals to the dashboard, and translates the driver's intent into mechanical action—to become a roadworthy vehicle. All three are essential, but their relative value is undergoing a dramatic redistribution.
For the past three years, the world's attention has been on the engine. OpenAI, Google, Anthropic, Meta—the brightest minds and the most capital have been poured into one thing: building bigger, more powerful engines. This is crucial; without an engine, nothing else matters.
But an industry consensus is forming: the curve of model capability improvement is slowing. Simply adding parameters and data is increasingly unlikely to yield significant performance breakthroughs; marginal returns are diminishing sharply.
Think of the automotive industry a century ago. While all manufacturers were competing on engine horsepower, Henry Ford realized something different: the key wasn't horsepower, but how to make that horsepower usable for ordinary people. So he invented the assembly line, standardized parts, and the entire system that transformed cars from toys for the elite into tools for the masses. Essentially, that was the Harness of the industrial age.
Today's AI industry stands at a similar inflection point.
As differentiation in engines narrows, the competitive focus naturally shifts to who can better utilize the engine. This mirrors how the arms race in CPU performance eventually gave way to competition over operating systems and application ecosystems. Chip performance matters, but real value is created in the superstructure.
In 2026, Anthropic released a set of engineering experiment data: the same model, the same prompt, running for 20 minutes in a simple setup cost $9 and produced completely ineffective core functionality; running for 6 hours with a full Harness cost $200 and delivered a fully playable game with all core interactions functional.
The model didn't change; what changed was the harness controlling it.
Everything outside the model—including code, configuration, execution logic, feedback loops, and constraint mechanisms—falls under the Harness category. The model is the source of capability; Harness turns that capability into a usable system.
We've had similar practical experiences within Tencent. With the same model capabilities, different scaffolding designs—such as what tools the model calls, how layered context engineering is done, how long-term memory is managed, and how workflows are designed—significantly impact actual usage effectiveness and token costs.
Harness Takes Center Stage
Why did Harness suddenly move from behind the scenes to center stage in 2026?
The fundamental reason is a shift in AI usage paradigms.
2025 was the year of the agent. The role of large models evolved from answering questions to executing tasks. Previously, you gave a model a question, and it gave you an answer—a one-time, clean interaction. But agents are different; they need to work continuously, remember context, call tools, decide what to do next, and self-correct when they make mistakes.
Just as apps defined the mobile internet era, agents are becoming the primary application vehicle of the AI era. Agents transform models from "answering" to "working." And work naturally requires a working environment: tools, file systems, feedback loops, and acceptance criteria. That system is the Harness.
Within this working environment, Skills are one of the most critical modules. Skills are text-described capability units that large models can directly understand, telling the model "what this tool is, what it can do, and how to call it." Their value lies in universality. Any Skills based on natural language descriptions can be called by different Harness frameworks. SkillHub is precisely the Skills circulation platform built for this era, enabling capabilities to be accumulated, reused, and shared across frameworks.
In engineering practice, the industry has made four key discoveries that directly drove this shift in focus.
Discovery 1: The Ceiling of Model Capability Lies Outside the Model, Not Inside It
The most compelling experiment comes from Nate B Jones: the same model, with only the Harness changed, saw programming success rates jump from 42% to 78%. LangChain's data corroborates this: the same model, with only the Harness altered, improved its Terminal Bench 2.0 score from 52.8% to 66.5%, moving from outside the top thirty to within the top five.
This overturns an assumption: to make AI stronger, you need to train better models. In fact, at this point, optimizing the "shell" outside the model may offer higher returns than waiting for the next generation of models.
Discovery 2: Constraints Guide Intelligence, Not Suppress It
The Cursor team, in large-scale agent experiments, found a counterintuitive phenomenon: when a model could generate anything, it wasted many tokens exploring dead ends; but when the Harness defined clear boundaries, the agent converged to correct answers faster. Constraining the solution space actually increased agent productivity.
Discovery 3: Harness Makes Large Models Safer
A large model without a Harness is like an intern without operating procedures—capable, but you never know what they'll do next. Harness, through permission boundaries, sandbox isolation, operation auditing, and manual approval nodes, confines the model's action space to a controllable range. Most Harnesses explicitly define: which systems can be accessed, which operations require secondary confirmation, and which data must never be touched.
This isn't a weakening of AI capability; it's a prerequisite for AI to truly enter enterprise production environments. AI must be trustworthy; only when it's used with confidence can it be affordable and truly usable.
Discovery 4: AI Cannot Reliably Evaluate Itself
Anthropic's engineers found that when an agent evaluated its own completed work, it confidently stated "done well," even when humans saw obvious quality issues. Their description: "Out-of-the-box Claude is a poor QA agent." This means that models alone cannot form an effective quality feedback loop; independent evaluation mechanisms must be established outside the model, which is one of Harness's core responsibilities.
These practical discoveries have converged into an irreversible trend.
Looking back at the evolution of AI engineering:
- 2022 to 2025: The keyword was Prompt Engineering—how to write a good instruction.
- 2025: Evolved to Context Engineering—how to dynamically construct the entire context.
- 2026: Harness Engineering arrived—how to build the entire working environment.
An analogy: Prompt Engineering gives the driver a map, Context Engineering gives them a navigation system, and Harness Engineering builds them a complete car with a dashboard, feedback systems, safety constraints, and autopilot.
Each evolution doesn't negate the previous one but encompasses it. Maps and navigation are important, but with only maps and navigation and no car, you're going nowhere.
The Driver's Awakening: More Powerful AI Demands More From Humans
In the engine-harness-driver triangle, the driver is the most easily overlooked role. The narrative protagonist for the past three years has been the model; 2026's new darling is the Harness. But what ultimately determines output quality is always the person in the driver's seat.
Here's a profound proposition: the more powerful AI becomes, the higher the demands on humans, not lower.
Consider autonomous driving. On the surface, it's about not having to drive. But someone who can safely supervise an autonomous system needs a deeper understanding of driving than an ordinary driver. They need to understand system boundaries, know when to trust the machine and when to take control, and make better judgments than the machine in emergencies. The driver in autonomous driving isn't a more relaxed role but a higher-order one.
The same applies to AI. Harnessing AI requires understanding both human engineering practices and AI's way of thinking, deep insight into system behavior, the ability to abstract error patterns into rules, and, crucially, the taste to build bridges between human wisdom and machine intelligence.
Taste. This word is being mentioned increasingly often. It's not aesthetic preference but something deeper: the ability to judge what is good, what is right, and what is worth doing. With the same engine and the same Harness, different drivers can produce vastly different results.
For most people, the Harness era is a more optimistic future.
Returning to the car metaphor. Today's automotive industry shows two seemingly contradictory trends: on one hand, the skill value of top drivers like F1 racers has never been higher; on the other, autonomous driving is making travel unprecedentedly safe and convenient for ordinary people. These trends aren't contradictory; they're two sides of the same coin.
AI is creating a two-tier structure. At the top, elite drivers—those who truly understand the engine and excel at designing harnesses—will produce the finest work. Their competitive barrier isn't execution but taste, judgment, and creativity. At the bottom, most people won't need to become high-level drivers to enjoy the capabilities AI brings.
This is the inclusivity of AI Harness. You don't need to understand every engine component or design every wire in the harness to benefit from AI-powered productivity tools. AI Harness is turning intelligence into infrastructure, making Intelligence as a Service a reality.
Put more intuitively: Harness is that "never-tiring digital employee." It doesn't take leave, doesn't forget things, and is on duty 24/7. More importantly, its capability boundaries are defined by you, its action scope controlled by you, and its output quality reviewed by you. Once configured, it can handle emails, organize data, generate reports, and run processes while you sleep. This isn't science fiction; it's happening today.
Enabling everyone to become an AI-enhanced "super individual," where one person can play multiple roles and get things done alone. This has always been our steadfast direction.
As the public becomes accustomed to higher-quality content output, their aesthetic thresholds and expectations naturally rise. They'll become more discerning, better at distinguishing genuine ingenuity from mediocre formulas. This, in turn, pushes professional creators to pursue more stylized, profound works. This isn't a zero-sum game but a mutually thriving ecosystem.
When Models Grow Their Own Hands and Feet
But this two-tier structure may not be the final state. We must recognize an ongoing trend.
As model capabilities continue to strengthen, with larger context windows, improved memory, and longer reasoning chains, models are growing their own hands and feet. Tool calling, context management, feedback loops, and memory systems that require external setup today are being internalized one by one.
The external scaffolding is thinning. In extreme terms, when models become sufficiently powerful, Harness might be entirely absorbed by the model. Like early cars needed complex external mechanisms to convert engine power, while modern electric vehicles have highly integrated engines and transmission systems, with simpler wiring harnesses because the engine itself "understands."
OpenClaw was the first lobster to "crawl ashore"; tomorrow might bring crabs, conches, mantis shrimp—different forms and focuses of Harness frameworks will continue to emerge and iterate. But these are surface phenomena; what's more important is that making large models grow hands and feet and actually work is an irreversible trend. Frameworks may change, but the paradigm won't revert.
When that day comes, the driver's role will upgrade from "operator" to "principal," no longer telling the AI how to run but where to go, and letting it find its own path.
But even if models absorb all tools and processes, one thing they can never generate themselves: the destination. Where to go, why go there, and how to judge if it's worth it upon arrival—these questions about direction, meaning, and value will always be human responsibilities. The stronger the model, the heavier this responsibility. Because when the machine can do anything, "what to do" becomes the only important question.
This confirms a simple truth: AI's value lies not in how powerful it is, but in how well we can harness that power to serve real scenarios, real people, and real needs.
Conclusion
Thousands of years ago, humans on the Eurasian steppes first put reins on horses. In that moment, human civilization gained unprecedented mobility: farming ranges expanded, trade distances extended, and ideas spread faster. What changed the world wasn't the horse's power, but the harnessing system humans invented.
Today, we stand at a similar juncture. Large models are this era's wild horses, astonishingly powerful. Harness is the reins we've invented, converting this raw power into controllable, predictable, collaborative capability. And the drivers—you, me, and everyone coexisting with AI—are the agents deciding where this power heads.
Artificial intelligence has officially entered the Harness era. The truly scarce capability isn't inside the model; it's outside. Taming a wild horse doesn't require a longer whip, but a well-fitted harness and a rider who knows the destination.
In this uncertain era, we must continue anchoring technological innovation, open collaboration, and value focus, making AI a truly usable, affordable, and trustworthy inclusive productivity tool.