The latest hot trend, Harness. Kai-Fu Lee and Qi Lu have invested heavily.

The biggest problems troubling shrimp farmers are, firstly, that Tokens are too costly, and secondly, that long - range tasks are unreliable.

If you assign OpenClaw to a long - range task that spans over 24 hours and leave it unattended in the middle, the results are often not optimistic.

Either the Tokens are used up when the task is only half - done, or after one step goes wrong, it keeps going wrong all the way, and the final output is completely useless.

The story of Summer Yue, the director of Meta's security alignment, having her email inbox emptied by OpenClaw overnight has also become a well - known anecdote.

https://x.com/summeryue0/status/2025774069124399363

Even earlier, there were cases where more than one Agent leaked sensitive company data to unauthorized employees.

All these incidents point to the same problem.

The smarter a single - entity intelligence is, the more likely it is to fail at critical nodes when placed in a longer time frame and a more complex collaboration chain.

This is like an intern with a high IQ but no supervision. They are amazing in the first half - hour but start creating problems for themselves in the second half - hour.

People in the industry are beginning to realize that just making the model smarter is not enough. There also needs to be something to manage how it uses its intelligence.

This something has a name that is being mentioned more and more frequently recently: Harness.

A New Consensus Emerging

The word "Harness" literally means horse gear.

In the context of AI, it refers to the control framework that connects the "model (the horse)" and "human needs (the rider)."

Put simply, apart from the "brain" of the Agent (such as Claude Opus 4.6), the rest is the Harness.

It doesn't participate in the task execution itself, nor does it try to become smarter. But it decides where the horse runs, how fast it runs, and when it should stop.

This judgment is not just the wishful thinking of one company.

In February this year, OpenAI published an official blog titled "Harness Engineering: Leveraging Codex in an Agent - First World," which used a set of experiments to prove that a three - person engineer team could make an Agent write a product with a million lines of code in five months through Harness Engineering.

https://openai.com/zh-Hans-CN/index/harness-engineering/

Anthropic also recently launched a new Agent architecture called Managed Agents, and its technical documentation repeatedly emphasizes the concept of "Agent Harness."

That is to say, after OpenClaw, the top players in the industry almost simultaneously discovered that Prompt Engineering and Context Engineering were no longer sufficient, and a higher - level constraint system was needed.

The logic of Harness may seem counter - intuitive at first, trading constraints for autonomy.

But the underlying principle is actually easy to understand. The more autonomy something has, the more likely it is to go astray. So, by putting a good enough harness on it, it can actually go further.

This consensus is rapidly converging.

For ordinary users, it's highly likely that a batch of new AI products will emerge in the next six months to a year. These products will no longer emphasize the size of their models but rather their ability to "tame" the models.

Right at this moment, a company that has been established for just over four months has quietly secured two rounds of substantial investment.

A Company Where Qi Lu and Kai - Fu Lee Rarely Appear Together

The company is called Nextie.

On April 13th, it announced that it had successfully completed two consecutive rounds of financing.

The angel round was co - led by Sinovation Ventures and Atypical Ventures, with Miracle Plus continuing to participate. Individual investors such as David Ku, the former global vice - president of Microsoft, also joined the investment.

David Ku

The capital reserve can support continuous innovation for the next three to five years.

Against the backdrop of a cooling primary market, this figure isn't particularly eye - catching.

What really makes industry insiders take a second look is its list of investors. Qi Lu and Kai - Fu Lee rarely appear together and jointly bet on the same Agent startup company.

Qi Lu

Kai - Fu Lee

The fact that two people regarded as the vane of the AI circle appear on the shareholder list of a company that has only been established for a quarter is a significant signal in itself.

However, what really gives weight to this story is not just the investors but also the person leading the team.

His name is Di Li, the former deputy dean of Microsoft Asia Internet Engineering Institute.

Di Li

In the AI circle, he has a more well - known identity: the father of Xiaoice.

The 6th - generation Xiaoice in 2018

To understand what Di Li is going to do this time, we first need to understand what he failed to achieve in recent years.

At the end of 2022, when the potential of the Transformer architecture was just starting to be fully recognized by the outside world, Di Li proposed within Xiaoice to purchase GPUs as soon as possible to accelerate the training of larger - scale models.

This proposal was shelved in the decision - making framework at that time.

That was a crucial window period when the capabilities of foundation models began to diverge rapidly, and Xiaoice missed this opportunity.

In February 2023, Di Li's team launched a project called X - CoTA.

Looking back now, what X - CoTA did was almost the same as the Chain of Thought (CoT) that later shocked the entire industry, which made the model "think for a while" before giving an answer and made the reasoning process explicit.

X - CoTA achieved an observable and traceable Chain of Thought construction with only about 2% of the parameters of GPT - 3.

However, it only lasted for a month.

In March of that year, it was shut down with the reason of "not understood, not allowed to continue."

By the end of 2023, Xiaoice's business in Japan was doing well, and there was money in the account. The team proposed to develop a reasoning model, but this direction was rejected again in the decision - making framework at that time.

Di Li later referred to this as "the only deep regret so far" in a public interview. In essence, he said that what was lost was not just a project but one and a half to two years of time.

Connecting these key points, we can see that there were actually several seeds in Xiaoice that could have blossomed.

Foundation models, Chain of Thought, and reasoning ability all precisely aligned with the subsequent waves of industry development.

It's just that these seeds failed to grow at critical moments.

On December 9th, 2025, Di Li founded Nextie with the core founding team of Microsoft Xiaoice.

Four months later, his non - compete period officially ended.

That is to say, he can finally start over and implement those ideas that he has repeatedly pondered over the years.

220 Years of Academic Literature

Creating Advanced Intelligence

What he is starting over with is called "Swarm Intelligence."

The idea of Swarm Intelligence didn't just emerge recently.

Di Li's team has sorted through the entire 220 - year human academic literature from 1800 to 2020, with only one goal: to figure out how human society, as a swarm intelligence system, gradually formed large - scale advanced cognitive collaboration.

This is the only verified sample of swarm intelligence that can continuously generate value so far.

The focus of this work is very specific.

Nextie's first product, launched in February this year, is called "Tuanzi" (tuanzi.ai).

After a user poses a question, dozens of Agents will "gather around a table," each approaching the problem from a different perspective, complementing each other, and engaging in debates. There are also professional processes such as voting and peer review in the middle.

Di Li internally calls this mechanism "cognitive collision."

There is a counter - intuitive aspect here.

According to common sense, multi - agent systems should consume more Tokens than single - entity systems, considering the complexity of multiple voices.

However, the data from Tuanzi shows that when achieving the same depth of thinking, the overall Token consumption is actually reduced by more than 50%.

The secret lies in the three words "coordination tax."

In traditional multi - agent architectures, as the number of steps and node branches increases, the context is carried along, copied, and passed down layer by layer, resulting in an exponential increase in Token consumption.

Nextie's approach is to "converge" at each step.

The purpose of actions such as debate, questioning, reflection, and voting is not to disperse information but to converge it at each layer before passing it down.

This is an experience borrowed from human society.

A well - run meeting is not just about letting everyone speak their minds but also about reaching a consensus after the discussion and moving on to the next topic with this consensus.

Tuanzi's Intelligent Depth Index (IDI) significantly outperforms single large - scale models, including GPT - 5.2 Thinking, in three scenarios: long - range multi - agent collaboration, high - difficulty research tasks, and large - scale population simulation.

However, all these are observed from a B - end perspective.

The truly interesting changes are hidden in the new product that Nextie will launch at the end of April.

Is "Xiaobingdao" Coming Back?

Capable of More

Di Li revealed in a recent interview that the team is fully committed to developing a new product similar in form to "Xiaobingdao."

Due to intellectual property restrictions, it won't be called by this name anymore, but the core concept remains the same, and it adopts a completely different new technology architecture.

The name "Xiaobingdao" holds a special place in the hearts of old users.

Its original design concept was to observe what kind of AI group is most suitable for each individual.

Some people need work support and emotional companionship, while others need more rational decision - making assistance. Everyone's "optimal AI lineup" should naturally be different.

When developing Xiaobingdao back then, there were many technical barriers.

Take a specific scenario as an example.

A user tells an AI on the island, "I just broke up with my partner."

With the technology at that time, this information was either broadcast crudely to all the AIs on the island, and everyone came to comfort the user, which made the user feel embarrassed;

or it was only known to that one AI, and the other AIs had no context at all, making it impossible to form a real emotional support network.

Neither way was ideal.

The problem doesn't lie in whether a single AI is smart but in whether there is a reasonable collaboration mechanism among these AIs to determine what information should be passed, to whom, and when.

This is exactly the problem that swarm intelligence aims to solve.