Guillermo Rauch: Raising Kids Who Think, and Why Trust Is the AI Moat

Vercel CEO Guillermo Rauch tells Delphi why trust is the moat after AGI, why agents are now his customers, and how he is raising six kids to think with AI. In this 55-minute conversation on the Delphi channel, the creator of Next.js walks through how Vercel won customers away from the cloud giants, why his team is moving its effort from developer experience to verification, which software he deletes on sight, and what the night his house burned down taught him.

TLDW

Rauch argues that startups win by building for the incoming wave instead of fighting the old one, and that Vercel’s long bet on headless, API-first architecture paid off when coding agents started building that way by default. He calls trust an output rather than an input: you earn it by sweating details, not by running tactics. With agents writing more of the code, he says the human job shifts to verification and capital allocation, and any tool that will not expose its data to agents should be deleted. He closes with his house fire and with how he is teaching his six children to build a solid foundation first and then use AI as a power-up.

Thoughts

Verification is the new developer experience, and that reframing is the most practical thing in the interview. For fifteen years Rauch’s whole brand was making it easy to tell a computer how to do something: React, Next.js, zero-config deploys. Now he says Vercel is shifting its focus from frameworks to “the checks and balances for the agent and the task.” The reason is visible in any repo an agent has touched. He describes finding a test whose only job was to assert that a constant equals itself, the kind of thing an agent produces by the thousand while the product it is guarding still breaks. His fix is the same one he tweeted years ago (“write tests, not too many, mostly integration”): test the product from its edges the way a human would use it, and only drop into the internals when that test fails. The point most people will miss is the second half. He refuses to “surrender understanding.” The end-to-end check tells you something is wrong; a human who understands the system still has to know why. Teams that let agents write both the code and the tests, with no one holding the model of the system, are building the next generation of the brittle software he complains about later.

His rule for software purchasing is blunt enough to be useful: if a tool refuses to expose its data and actions to agents, delete it “like the plague.” He compares it to a company that opted out of the www in the nineties, and he dismisses the data-moat defense with “I genuinely don’t care.” The underlying claim is that the coding agent is to AGI what the web browser was to the internet, the universal client, and that MCP is becoming its HTTP. What makes this more than a slogan is the next layer he adds. Shipping an MCP server is not enough. Agents hit an API over and over, so clunky rate limits and inconsistent data break the loop, and he says much of modern software runs on “APIs that lie,” patched over for years by human engineers writing defensive hacks. Agents do not tolerate that, and they call at a million times the rate a human would. That is a quiet warning to every SaaS company with a decent API on paper: the bar is no longer “has an API,” it is “holds up under agentic pressure.”

The capital allocation idea follows from both. If agents do the work and verification loops check it, the human decision becomes where to spend tokens, compute and time. Rauch’s example is a performance loop he ran himself, where the agent came back and said, in effect, we are fast now, the next step costs another thousand hours, do you really want it? That is a budgeting question, not an engineering one. He predicts a lot of companies that shipped flashy demos for likes will soon be reallocating heavily into security loops. The framing is useful for anyone running agents today: your job is less “how should this be built” and more “is this worth what it costs, compared to everything else the agent could be doing.” He also notes he has invested in companies that hand an agent a seed round and let it run the business, and says some have shown success at small scale, which is either the logical end of this idea or a sign of how fast it is moving.

His parenting answer is the part of the video title most people will click for, and it is more conservative than you might expect from someone who sells agent infrastructure. He does not want his kids handing their thinking to AI. He wants a solid foundation first (one son draws beautifully with a pencil, another may need hardcore systems engineering to build mission control software) and then AI on top as a “power-up,” the way the personal computer and Google expanded his own intellect like extra cache layers on memory. He also spends a lot of time helping each child find their ikigai, the overlap of what they love, what they are good at, and what the world values. With six children he says he has “sampled through a probability distribution” of personalities, and the lesson is that they will not all be good at the same things. Five years ago he would have taught them all JavaScript. Now he would teach the tools, and the fundamentals only where the child’s interest runs deep.

Trust as an output, not an input, is the line from the description that will get quoted, and the gym metaphor that goes with it is a good one. Abs are made in the kitchen, not by crunches; a brand is made by security, docs, fixing bugs and simplicity, not by meetups and tweets. His model of how customers decide is the strongest part: people sample five or seven scattered interactions with a company, never write them down, and one day form a gut verdict when they need hosting. That is why he says you have to be excellent everywhere every day, and why he keeps a Slack channel literally called “scream” on high notification. It also explains how he closes the interview. Asked what conversation he hopes people have with his Delphi, he says you come for the outputs and stay to learn how the sausage is made. Trust, in his telling, is just the visible sum of a lot of invisible inputs.

Key Takeaways

Rauch never thought of Vercel as competing with AWS. His meta advice for startups is to position for the incoming wave of technology rather than fight the one that is losing steam.
Delphi, a 12-person team, moved from AWS to Vercel during a codebase rewrite and no longer needs a full-time infrastructure engineer.
Instagram’s tiny team beat Facebook partly because Amazon S3 let them store every photo with an API call. Rauch hopes Vercel will one day host the single-person trillion-dollar company.
Enterprises copy startups like a midlife crisis, so focusing on innovators produces a “trickle-up effect,” sometimes literally when a big company acquires a startup built on Vercel.
Pitching Vercel as “AWS but better” was naive. Winning in e-commerce, where a faster site means more money overnight, gave Vercel a clear ROI story and a niche to expand from.
Grow by concentric circles and natural adjacencies, like Tesla’s master plan. Jumping into unexpected adjacencies risks the company’s identity and its brand permission.
Trust is an output. It comes from security, documentation, bug fixing and simplicity, and customers build it from a handful of scattered interactions they never consciously record.
Most founders are not confronting enough pain. Read the thousand churn surveys and call the customers who rage quit.
Companies and products should mirror each other, which is why Rauch watches an internal Slack channel called “scream about internal developer experience” on high notification.
Agents are now a third kind of Vercel customer. Rauch wants agents to pick the best product honestly, and warns against both vendors and model labs gaming that choice.
The autonomy paradox: the best agent experience is one where you were not involved, which means the agent made a thousand decisions for you. It must still learn your preferences and push back.
Vercel is shifting focus from developer experience to verification and evaluation, testing from the edges like a human user while keeping people who understand the system.
Coding agents plus Vercel let the company delete four or five overlapping SaaS vendors. Rebuilding bigger systems still means owning security, durability and long-term maintenance.
Delete tools that refuse to work with agents. MCP is “the new HTTP,” and APIs must also survive the load and consistency demands of agent loops.
For his six kids: help each find their ikigai, build a solid foundation, then teach AI tools as power-ups for their intellect and creativity.

Chapters

0:39 How Vercel Competes With a Giant Like AWS

The host opens with Delphi’s own story: the team moved from AWS to Vercel while he was out of office, and it saved them time and money. Rauch says every product he builds starts from personal pain, in this case feeling stupid the first time he tried the cloud after a decade as an engineer. He frames Vercel as “self-driving,” now “agentic,” infrastructure, and says old technologies never disappear, they just lose attention. Startups should ride the incoming wave, and he points to Instagram on S3 as the model for tiny teams doing the work of thousands.

10:10 Why “AWS but Better” Was Naive

Enterprises imitate startups, and Rauch spent years pushing headless architecture, partly so his technology had somewhere to plug in. Agents have now persuaded the world to build that way. Early Vercel tried to be all of AWS but better; e-commerce gave it a niche where speed equals revenue and the value is easy to explain. He demos the new vercel.store with virtual try-on to high schoolers, then cites the Tesla master plan as a model for growing from a niche through concentric circles without losing brand permission.

17:15 Trust Is the Moat After AGI

Asked about Vercel’s cult-like developer loyalty, Rauch says trust will be one of the most valuable moats for a post-AGI business, but it is an output of the work. He uses the fitness line that abs are made in the kitchen: crunches are the tactics, nutrition and sleep are the foundation. He credits Braun’s influence on Apple as the inspiration for bringing design craft to developer tools, when caring about a README or an open source logo was a wild idea.

24:24 You’re Not Confronting Enough Pain

Building a company is a daily workout, and Rauch says many founders show up without pushing, like the guy on his phone at the gym. Real effort means reading every churn survey and contacting customers who left. Internally, he engineers the employee experience as carefully as the customer one, reading a Slack channel called “scream” so the company mirrors the product it sells.

27:46 Your Next Customer Is an Agent

Vercel’s customers went from developers, to non-technical builders through v0, to agents. Rauch sees one spectrum of “builders” with a low floor and a high ceiling. On why AI tools keep picking Vercel, he says agents should pick the best product, and nobody, including model labs, should hack that. He describes an autonomy paradox: the best agent experience means you were absent for a thousand decisions, so alignment with your intent matters more, not less.

33:45 AI Writes 2 Million Tests, Then Ships Garbage

Vercel is moving its focus from frameworks to verification. Rauch’s old tweet “write tests, not too many, mostly integration” riffs on Michael Pollan’s line from In Defense of Food, and he says agents writing millions of trivial unit tests have proved his point. Test the product like a human would, and do not surrender understanding of how the system works.

37:05 The Future Is Just Capital Allocation

Humans will decide where tokens, compute and time go, with security the current priority. Rauch describes an agent telling him further performance gains would take a thousand more hours, and people wiring agents to wake them at night for hard tradeoffs. He has invested in fully autonomous companies built on Vercel, where an agent gets a seed round and runs growth, marketing, sales and engineering. On build versus buy, he says rebuilding big systems means owning the full maintenance burden.

43:19 Delete Agent-Hostile Tools, MCP Is the New HTTP

Coding agents let Vercel replace four or five low-common-denominator SaaS tools. Tools that will not speak agent get deleted, data moat or not. MCP is the new HTTP, but performance, rate limits and data consistency matter too, because much of the internet runs on APIs that lie. Agents call APIs at a million times human rates, and Vercel’s Fluid compute is built to scale from zero to many regions to meet them.

47:30 The Night His House Caught Fire

His darkest moment came while building his first startup alone in San Francisco, when a neighbor’s cigarette set the house on fire in the middle of the night. He lost his possessions, two people were badly hurt, and he had to start over materially and spiritually. It pushed him into meditation and an appreciation of every day.

49:50 Raising Six Sub-Agents: Teaching Kids to Think

Rauch jokes that his six kids are sub-agents. He helps each discover their ikigai, builds their foundations, and then teaches AI tools as power-ups, the way Google extended his own memory. One son draws with a pencil better than he ever could; Rauch wants him to also build his own creative tools. He ends by saying he will give his Delphi to his kids, and that what matters most is how you live, think and work on yourself.

Notable Quotes

“One person plus agents can now do the work of hundreds of people, thousands of people”
Guillermo Rauch, on hoping Vercel hosts a single-person trillion-dollar company

“I think trust is an output, not an input.”
Guillermo Rauch, on how Vercel earned developer loyalty

“you’re not confronting as much pain as you could be about the things that are broken or shitty.”
Guillermo Rauch, on founders who show up but do not push

“I also don’t want to surrender understanding.”
Guillermo Rauch, on verification loops for agent-written code

“in terms of what the web browser did for the internet the coding agent is doing for AGI.”
Guillermo Rauch, on why coding agents replace low-code SaaS tools

“MCP is almost like the new HTTP.”
Guillermo Rauch, on tools that refuse to speak agent

“one of the gifts of having six children is that I’ve now sampled through a probability distribution of all kinds of different personalities”
Guillermo Rauch, on raising kids to think in the AI era

Watch the full conversation here.