Thursday, June 25, 2026

There are two things worth talking about today, but they are actually two sides of the same thing: the Agent is running, but his seat belt is not fastened yet.

To put it bluntly, the bloody and tearful post on Reddit only said one sentence-if you let LLM, which has no authority boundaries, access the database, it will give you a big job sooner or later. But the problem is that the bean bag professional version launched an agent-driven office model today, and the 200 million daily applications have begun to make Agents work hard. Cursor is working on a fully autonomous training model. The ponytail project on GitHub makes Agents "think like the laziest senior engineer," headroom is doing tool output compression to save tokens, and codebase-memory-mcp indexes the code base into a knowledge graph. The entire ecosystem is rushing in the direction of "Agents do real work."

But there is a fundamental contradiction here: the ability boundaries and privilege management of agents are far from keeping up with the growth of capabilities. You ask an Agent that can tune APIs and read and write files to work. If it does a crazy day, delete tables and change configurations, you won't even have time to cry. If Doubao dares to go into office mode, there is a high probability that he has tight authority-scenarios such as research reports and financial report analysis cannot touch the production system. But once you go down and the Agent wants to operate the real system, this is a ticking time bomb. My judgment is very clear: Agents can be used freely in read-only and generate scenarios, but whenever writing operations are involved, there must be hard isolation, so don't save this effort.

The chip line is also worth talking about. Huang Renxun said that the AI infrastructure cycle takes decades, but he is not the only family at the poker table. Qualcomm announced that it will launch third-generation AI chips in 2027, targeting data center revenue of US$5 billion. Microsoft Azure has already said it will deploy them. OpenAI's first self-developed chip was unveiled by Broadcom. Micron and Anthropic collaborated to get Claude to participate in chip design-AI in turn designed AI chips, closing the loop. Nvidia's CUDA ecosystem is indeed difficult to replace, but when major customers such as Microsoft Meta are engaged in self-research or seeking cooperation with Qualcomm, the bargaining power will be loosened sooner or later. Qualcomm is cutting from mobile phone chips to data centers. The goal of US$5 billion in 2027 seems radical, but with Microsoft's endorsement, it is not a pie.

These two lines are actually the same thing: the computing power at the bottom is diversified, the Agent at the top is exploding, and the middle layer-how to safely and efficiently connect the Agent to the real system-is now a huge blank area. Whoever can provide reliable authority isolation, context management, and error recovery at this intermediate level will be valuable. Headroom and codebase-memory-mcp are in the right direction, but not enough. Ordinary developers now get started with Agent, first practice from read-only scenes, and don't let it touch the production library as soon as it comes up, otherwise you will be the protagonist of the next Reddit bloody and tearful post.

When senior Reddit engineers complained bitterly that AI blew up production databases, and a large number of open source tools have emerged on GitHub that restrict AI from scribbling code and compress redundant tokens, we are witnessing the turning point of AI Agents from a technical carnival to engineering disillusionment. Doubao has launched a professional version of the Office Agent, and WeChat AI has maintained extreme restraint in core payment and social functions, both confirming the same underlying trend: the cleverness of the big model has overflowed, but the controllability is seriously lacking. Half a year later, the competitive axis of AI applications will completely shift from parameter scale to engineering binding. Middleware and flow control platforms that can provide certainty and are good at putting engineering shackles on AI will replace shell applications as the new core infrastructure. Teams that blindly believe in AI autonomy will face the risk of catastrophic production accidents.

Echoing the tightening of upper-level applications, the bottom-level computing infrastructure is staging a secret battle to break monopoly. Huang Renxun shouted that the AI infrastructure cycle lasted for decades, but Qualcomm made a high-profile move into the data center and OpenAI teamed up with Broadcom to launch the first customized chip, revealing the technology giant's urgent desire to survive to get rid of its dependence on a single computing power. This is not only a difference in hardware routes, but also a reconstruction of computing power pricing power. A year from now, the high premium for general-purpose GPUs will be severely diluted, and custom architectures for long text reasoning or multimodal generation will reshape the cost structure of the entire industry.

This extreme squeeze between the upper and lower ends is accelerating the shuffle. The upper layer needs extreme certainty, and the lower layer pursues extreme customization. The general shell entrepreneurs and traditional IT positions caught in the middle will bear the brunt. Oracle's massive layoffs due to AI replacement are just the prelude. The winners in the future belong to those pragmatic who can establish tight guardrails at the application level and grasp the right to define vertical computing power at the bottom level.

Huawei's smart driving sounds like reassurance to consumers, but if you look at the terms carefully, it is more like passing on insurance costs to car owners. The announcement states that protection will only be provided if accidents occur on legally driving roads and areas. What is a legal road? Do construction sections and temporary diversions count? Who defines it? Moreover, this price is tied to the price increase. The high-end feature package has been restored from 32,000 to 36,000, and the terminal landing price has increased from 12,000 to 15,000. Huawei is raising prices while talking about the bottom line. It is essentially asking you to pay more to buy an invisible insurance. The claim threshold and exemption clauses of this insurance are probably more complicated than any commercial insurance on the market. Historically, it is not the first time that a car company has taken responsibility for smart driving. Tesla also talked about "fully autonomous driving" back then. As a result, after the accident, the lawyer's letter was faster than the system iteration, and it was ultimately attributed to "the driver's failure to take over in time." I am afraid that we will not know how much Huawei can cover in judicial practice until the verdict of the first major accident is issued.

The excitement in the AI circle today is reflected in a bloody and tearful post from Reddit. The data engineer's warning was very direct: Don't let AI touch the production environment. You think you're a senior expert, but you're actually a master of warehouse explosion. On the same day, the professional version of Doubao was launched, which mainly promoted an agent-driven office task model. It claimed that 200 million daily users could use AI that could work. On the one hand, there are painful lessons from the real world, and on the other hand, there is the rapid product release of major manufacturers. This sense of division makes people have to be vigilant. Agent has been packaged into a next-generation productivity tool with "autonomous decision-making and automatic execution" from concept to implementation in the past two years, but its reliability has not been tested in a large-scale production environment. The Reddit post is not an isolated example. If you just look through GitHub, you can see a bunch of issues that use LLM to operate databases and operate servers causing disasters. Nowadays, Doubao, WeChat AI, and even Enterprise WeChat have begun to internally test AI Agents, handing over this uncontrollable autonomy to hundreds of millions of users. Once a chain failure occurs, what will be blown up will not only the database, but may also be the credibility of the entire ecosystem.

The bigger bubble is at the capital level. On the same day, WSJ and The Information exposed the collapse of Ultraman's "technology saint" persona. He held 0 shares but used OpenAI as a private cash machine. Confidential IPO documents also exploded a $665 billion bomb. Huang Renxun made a fuss at the shareholders 'meeting, saying that the AI infrastructure cycle lasts for decades, and the more computing power, the more revenue, the more revenue. However, the Nasdaq index has begun to pull back, chip stocks have continued to fall, and Hertz car rental, which is related to AI, has fallen by 41%. OpenAI's valuation model is based on the assumption of unlimited future growth, but the reality is that downstream applications of the big model have not yet found a sustainable profit model. Although the Bean Bag Professional Edition provides the 2.1 Pro model for free, it is behind the old path of burning money for the market. Ctrip's quarterly profit fell 41.57% year-on-year, indicating that consumers have already felt the chill of high energy prices and geopolitics, while AI's infrastructure investment is accelerating. If this scissors gap continues, the first one who cannot withstand it will not be Huang Renxun's vision, but those individual investors and institutions attracted by high valuations.