A partner at HSG stated: From just being able to talk to being able to act, AGI has already arrived in 2026.

While the world is still debating "what is Artificial General Intelligence (AGI)", Sequoia Capital, the oldest venture capital firm in Silicon Valley, has grown impatient.

At the beginning of 2026, Sequoia partners Pat Grady and Sonya Huang co - authored an article titled "2026: This is AGI". They directly asserted: Stop waiting, it's already here.

It's not the awakening of Skynet, nor the rule of robots over humans. Instead, it quietly hides in those "long - cycle agents" that can work continuously for dozens of minutes or even hours.

"Get ready," they wrote, "your vision for 2030 has already been realized in 2026."

They didn't just say this casually. They gave a solid example: An agent helped a founder precisely identify a nearly perfectly - matched recruitment target in just 31 minutes.

From "having a chat" to "getting things done", the boundaries of AI's capabilities are being redrawn.

01 Defining AGI? Being able to "figure things out" is enough

The academic and industrial circles have been arguing about the definition of AGI for many years.

Grady and Sonya Huang recalled that in the early years, when they asked top researchers how to define AGI, the researchers often looked at each other blankly and finally blurted out: "We each have our own definitions, but we'll know it when we see it."

It sounds rather mysterious. But now, these two investors have decided to put aside philosophical debates and give an extremely practical, even somewhat "crude" definition:

"AGI is the ability to figure things out. That's all."

They explained that a person who can "figure things out" needs knowledge, reasoning ability, and the ability to iterate and learn from mistakes. The same goes for AI: Knowledge comes from pre - training (like ChatGPT), reasoning relies on stronger computing power (like the o1 model), and iteration and learning from mistakes depend on "long - cycle agents" (like the recently emerged Claude Code).

When these three elements come together, an "agent" that can work autonomously, correct itself, and know what to do next without instructions appears.

To put it simply, they don't care how complex the inside of AI is. They only care about one thing: "What on earth can this thing do?"

For them, an AI that can truly solve problems and influence the real world is "general - purpose".

02 31 - minute combat: From vague instructions to a precise candidate, the AI headhunter is on the job

The definition is abstract, but the application scenarios are starting to take shape.

Grady and Sonya Huang described a specific task: A founder needed to find a head of developer relations, with the requirements of "being good at technology, respected by engineers, and really loving to use Twitter".

The instructions were vague, but the AI agent took on the task. Instead of conducting a simple keyword search, it carried out a set of complex operations like a senior headhunter:

Not just looking at resumes: It first searched for relevant positions in competing companies (Datadog, Temporal, Langchain) on LinkedIn but quickly realized that job titles were unreliable.

Turning to assessing capabilities: It went to YouTube to dig for industry conference speeches and specifically targeted speakers with high audience interaction - only those who can speak well have real skills.

Digging out real personalities: It took the list of speakers it found and "investigated" them on Twitter. It screened out the "tool people" who only retweeted company posts and locked in real people with real fans, who dared to express their opinions, and whose posts had an "online feel".

Catching signs of job - hopping: It checked which people's posting frequency had decreased in the past three months, which might mean they were dissatisfied with their current jobs.

Cross - verification and elimination: After locking in several targets, it eliminated those who had just been promoted or started their own businesses and finally focused on one person - a senior professional who had just been laid off by their company, highly matched in the technical field, and hadn't updated their LinkedIn profile for two months.

The agent even came up with a plan: Not only did it find the person, but it also drafted a recruitment email with a sincere tone and a precise angle of approach.

The whole process only took 31 minutes. Instead of receiving hundreds of resumes, the founder got a specific target with a high probability of success and an action draft.

Grady and Sonya Huang summarized that this is "figuring things out", autonomously exploring, making mistakes, and changing directions within a vague goal until a path is cleared.

"The agent has the skills of an excellent recruiter, but it never gets tired and doesn't need to be told specific methods."

Of course, AI still makes mistakes and has "hallucinations". But the trend is irreversible: It is changing from an "intern" that needs to be taught step - by - step to a "prospective colleague" who can independently take charge of a task.

03 Core breakthrough: Long - cycle agents that enable AI to "think longer"

Why can today's AI "run" a 31 - minute task in one go?

The key lies in the "long - cycle agents", which can be understood as adding "sustained focus" and "task management" add - ons to the AI's brain.

Previously, AI models could only perform one inference in a few seconds. Now, there are mainly two ways to make it "last" longer:

Reinforcement learning: Repeatedly "training" the model during the training process to make it learn to stay focused on long - term tasks. This is the main battlefield for top AI labs.

Agent frameworks: Designing external auxiliary tools for AI to help it manage memory and plan steps. This is where application - layer companies (such as the teams behind Claude Code and Manus) are focusing their efforts.

The progress is exponential. Data tracked by the independent non - profit evaluation agency METR shows that the ability of AI to complete long - cycle tasks doubles approximately every 7 months.

Calculated at this speed:

By 2028, AI will be able to reliably complete a full - day's work of a human expert.

By 2034, it will be able to finish a human expert's one - year workload.

By 2037, it will be able to handle tasks that humans need 100 years to complete.

What does 100 years of workload mean? It could be analyzing all historical clinical trial data, digging out patterns from a vast amount of customer service records, or completely rewriting the complex tax laws.

04 The era of "hiring" AI colleagues has begun

A direct litmus test for AGI is: Can you "hire" it?

Grady and Sonya Huang believe that soon you will be able to "hire" professional AI agents to work for you, just like hiring human employees. Currently, you can already "hire" models like GPT - 5.2 and Claude, and specialized "AI employees" are emerging:

Your "AI specialist doctor" (Deep Consult from OpenEvidence)

Your "AI legal assistant" (Harvey agent)

Your "AI network security officer" (XBOW agent)

Your "AI chip designer" (Ricursive agent)

Even your "AI researcher" (GPT - 5.2)

This means that in 2023 - 2024, AI was mainly a "talker", that is, a smart chatting partner. In 2026 - 2027, AI will become an "executor", just like a real colleague.

Our work mode will be disrupted: Instead of asking a few questions a day, we will have several AIs working for us simultaneously every day.

Our roles will also change: From being a hands - on "executor", we will become a "manager" of an AI team.

"All discussions about'selling work' are now possible."

Conclusion: Stop just chatting and assign tasks to it

The exponential growth of long - cycle agents has already started. Today, it can run reliably for 30 minutes; soon, it will be able to take over your full - day work; in the future, its vision will be measured in "centuries".

When AI can not only answer "what is it" but also continuously explore "how to do it", what we are facing is not just a tool upgrade but a reconstruction of the productive relations.

Grady and Sonya Huang finally said: Stop treating AI just as a chatbot or a search engine. It's time to assign tasks to it.

In 2026, AGI may not arrive in a spectacular way, but it has transformed into countless agents in the digital world that quietly "figure things out" and take action.

This silent revolution has already begun.

(Jin Lu, a special translator for Tencent Technology, also contributed to this article)

This article is from "Tencent Technology", author: Boyang. It is published by 36Kr with authorization.