跪拜 Guibai
← Back to the summary

Why Every AI Lab Is Betting on Coding Agents to Reach AGI

Yesterday, many people asked in the backend, so let me explain a point.

Why is Coding Agent a very important link on the way to AGI?

The Codex, Claude Code, Kimi Code, Grok, and so on that we are currently using are essentially Coding Agents. Almost every global AI company now treats this as one of its most important areas.

The reason is simple: The coding field has several unique advantages.

  1. Automation of the digital world. All the information and operations that human elites have accumulated on computers over the years, as long as you want to automate them, programming is indispensable. In the AI era, automation comes first, then intelligence.

You can see now that almost everything you do on a computer can be handled by AI.

  1. Open-source corpus. Many people today are touching GitHub for the first time in the AI era, but it has existed for over a decade.

I once heard a joke: no one would make a money-making project or product freely open source, unless they are a programmer.

It can be said that programmers are probably the most open and open-source community. And mature projects and source code in software programming are the best and lowest-cost training corpus for AI.

  1. Fast verifiability in software engineering. One of the biggest difficulties in the AI era is how you judge whether the result given by AI is correct or wrong.

Some fields cannot do this, or the cycle is long. For example, medicine: if AI gives you a plan or makes a drug, the verification cycle alone is very difficult.

But software engineering is different: whether the thing you make has bugs can be quickly verified once it is done.

These are all unique advantages for AI.

However, there have been detours. At the beginning, all top companies wanted to take the ultimate route to AGI. Anthropic walked this path first and succeeded. Now everyone has realized that AGI cannot be achieved in one step; it needs to be taken step by step. Coding Agent is a subset of AGI, but it is one of the most important links, and I even think it can be considered the basic infrastructure of AGI.

So Anthropic walked this path relatively early and currently focuses only on this path, but OpenAI, Google, and other companies have never given up research on the AGI route while catching up.

Looking at Google's current AI product line, it covers almost everything from coding, images, videos, music, and even world models.

Yesterday I complained on X that Google's Antigravity has the worst user experience and design among all agents I have used, but the Gemini 4 Carbon new model does feel quite good.

This also includes the domestic DeepSeek Agent. I really think its experience is not good, but that only represents the present.

AI Agents only truly appeared this year. For the architecture and design of such agent products, you can consider that there is no prior accumulation or experience at all; they all start from zero and build up step by step, so everything takes time.

Do not judge every company's future by the current standards. Look at relative progress. At least in the past two months, I think Google has been the fastest in relative progress.