Vercel CEO reveals his AI agent coding workflow: 4 habits to adopt now, stop blind auto-merging

Guillermo Rauch, CEO of Vercel — the company behind Next.js — showed on his own machine how he works with AI agents. According to a video by the TechArk page, his sharpest line is: telling an agent "make no mistakes" and then auto-merging is building a pile of trash. He types a single letter and the agent opens the research folder, searches, and records on its own — and he has four habits any developer can adopt right away.

Quick answer: According to the TechArk video, Vercel CEO Guillermo Rauch works with AI agents through four habits: one research folder + an agents.md file + a one-letter alias to call the agent; not reading every line of code but understanding the code the agent writes; writing rules into agents.md instead of praying; and letting the agent try to break new infrastructure to measure latency and learn lessons. He warns that speed without direction is just PR spam, and that the opposite of a 4-hour week is happening: developers must put in more effort to prompt and to verify.

Agentic engineering is a way of building software in which developers hand work to AI agents that execute on their own — searching, writing code, testing — while humans keep the roles of setting direction, writing rules, and verifying results. The agents.md file is where those rules are written for the agent to read and follow in the project.

What does Guillermo Rauch say about auto-merge and "make no mistakes"?

According to the video, Rauch says it plainly: telling an agent "make no mistakes" and then auto-merging is building a pile of trash. A vague instruction does not replace concrete rules, and merging code without understanding it only accumulates risk.

He also warns that those still hesitating to work with AI agents will be lapped by other teams — but fast is not the same as winning. Speed needs direction; otherwise you are just spamming PRs, like driving at full throttle without knowing the road: the faster you go, the more lost you get.

How does one letter "R" replace Google?

In the video, Rauch opens his machine to show: he types a single letter "R" and the agent opens the research folder, searches, and records the results by itself. For him, it replaces Google altogether.

This is the first habit packaged: one research folder, one agents.md file stating the rules — what to search, where to save — and a one-letter alias to call the agent. Ask, then go do other work; when trying a new technology such as Next.js, create a new research folder right away.

What are the Vercel CEO's four habits for coding with AI agents?

Habit	How it works (per the video)
1. Folder + agents.md + one-letter alias	One research folder, rules in agents.md (what to search, where to save), call the agent with one letter; ask and go do other work
2. Don't read every line, but understand the code	Walk around the code, then redraw the diagram of which function calls which; refuse not to understand the code the agent wrote
3. Write rules into agents.md instead of praying	Write concrete rules for the agent, e.g. requiring it to message him immediately when an important event happens in the program
4. Let the agent try to break new infrastructure	Let the agent find ways to break it, measure latency, draw lessons — but never auto-merge the parts vital to the system, like security

The second habit gets the most emphasis: no need to read every line of code, but you must understand the code. He praises the approach of walking around the code and redrawing the diagram of which function calls which, and closes with a refusal not to understand — he disagrees with anyone who says there is no need to read code anymore. The person who understands agent-written code, in his view, is worth more than the person who just presses auto-merge.

The third habit turns wishes into written rules: instead of praying the agent does the right thing, write it into agents.md. According to the video, a single line of rules like that is more truly agentic than working by luck.

When should you accept technical debt, and when not?

Rauch draws a clear line between two cases. Where speed is needed, accept technical debt — trading speed for deliberate imperfection. But for vital parts such as security, never let it be automatic: those places require a human who understands and verifies.

That is also the boundary of the fourth habit: let the agent try to break new infrastructure to learn, but never hand automatic merging to the parts that decide the system's safety.

Why doesn't AI bring a 4-hour work week?

Here is the video's twist. Those dreaming of a 4-hour week thanks to AI are told the opposite is happening: AI grinds its users — developers will have to put in more effort to prompt and to verify.

In other words, agents do the manual work faster, but the human work — setting direction, writing rules, understanding code, verifying — gets heavier, not lighter. Higher productivity does not mean more free time if the quality bar rises along with it.

The video ends with a small challenge for tonight: create a research folder, write an agents.md file, set a one-letter alias — and ask yourself whether you are the person who understands agent-written code, or the person pressing auto-merge.

Frequently asked questions

1. Who is Guillermo Rauch? He is the CEO of Vercel, the company behind Next.js, and the person sharing his AI agent coding workflow in the video posted by TechArk.

2. What is the agents.md file for? It holds the rules for the AI agent in the project: what to search, where to save, when to alert a human — instead of hoping the agent guesses the intent.

3. What does "one letter R replaces Google" mean? Rauch types one letter and the agent opens the research folder, searches, and records results by itself — the lookup workflow is automated instead of manually searching Google.

4. Why shouldn't you auto-merge agent code? Because, according to Rauch, merging code without understanding it only builds a pile of trash; and vital parts such as security must never be automatic.

5. Will AI bring a 4-hour work week? According to Rauch, the opposite: agents work faster, but humans must put in more effort to prompt and to verify.

This article was rewritten from a video by the TechArk page posted on 14 November 2026, based on Guillermo Rauch's sharing and the video's automatic transcript. The statements and habits in the article follow the video's content; some unclear passages in the transcript have been paraphrased at a general level.