Coding agents now drive 60% of Vercel deployments, up from 3%

1. Decade versus ten months — ▶ 8:46

   Vercel took about 10 years to reach 1 billion deployments, then added another 1.5 billion in the next 10 months. The driver is coding agents, which write the code but need somewhere to ship it: "code doesn't do anything by its own." Vercel has become the place agents plug into, and OpenAI and Meta deploy there. It also distributes tokens and offers every model, closed or open, so it sees agent activity and model demand together. Rauch calls that "an index that tells you where AI is going."

2. Reject non-understanding — ▶ 11:47

   Two schools exist: tell the agent anything and press ship, or use AI and understand what it did. Rauch takes the second and recommends it to his children. He reads code and asks the AI to explain it back, acting like an engineering manager who still loves code. Over the weekend he shipped a proof system meant to stop agents shipping slop or leaking data. His split: vibe coding has a place, but capital-S Systems, meaning critical infrastructure and national security, need verification. Formula 1 cars and hospitals already run on code.

3. From 3% to 60% — ▶ 14:53

   At the start of the year 3% of Vercel deployments came from coding agents. Now 60% are agentic: "robots telling other robots what to do." How much of that 60% a human actually understands is "hard to say." Vercel became the agents' favorite because it open-sourced everything for 11 years, and models trained on that material work better with its rules. The root is his own history: learning English and hunting for free tools in Argentina led him to open source.

4. Websites for agents — ▶ 18:22

   The web was machine-to-human, and the next era is agent-to-agent. An agent visiting vercel.com gets no animated gradient, only precise information about the business through a Vercel Cloud capability. Rauch says every business must retool or be "left out of the most exciting economic opportunity of our times." Agents already do the building: asked to make a website, an agent visits Vercel for instructions. A DoorDash agent built on Vercel lets people order a burrito by text. Much of the internet was built to block bots, and Vercel takes the opposite posture.

5. Bot intelligence, agent buyers — ▶ 21:15

   John Collison's argument: if everything is a bot, bot detection is pointless, and what matters is bot intelligence, meaning whether the bot is having a good experience. Rauch applies this to ChatGPT, Grok, Claude and search crawlers hitting his site. Vercel partners with Stripe and Shopify so agents can transact. An agent can buy Browserbase through the Vercel marketplace with no human intervention. Models have training cutoffs and rely on search, so businesses should publish transcripts and content agents can find.

6. Protocols replace HTML — ▶ 24:54

   HTML served the human internet, and agents need protocols. MCP lists the tools a service exposes, and Vercel's covers deploying software and listing projects. The machine payment protocol and universal commerce protocol let agents buy. Rauch gets account briefings from a Slack agent that queries Salesforce, so he never opens it. Interfaces become generative: a profile for a person, context for a company, tuned to experience level, rather than layouts fixed by a product design team. He expects every product and service to be transformed this way within a few years.

7. Yes code, not no-code — ▶ 30:08

   Rauch takes the opposite stance to software aggregation. He believes in personalized software generated by coding agents for one user or one enterprise domain, where ROI is far greater. He calls Vercel a yes-code company, the opposite of no-code, because agents make code free. Buy Vercel as a platform and "you never have to buy software again." Salesforce and SAP survive as systems of record, critical data infrastructure, because software needs data. Headless means the agent fetches the data instead of people visiting the interface.

8. Soldiers without frontal lobes — ▶ 35:25

   Autonomy is what changed since ChatGPT: put the model in a loop and train it to run longer. Agents will do whatever it takes, like soldiers with the frontal lobe removed. Asked to find an email, one might spend 40 minutes searching, including Crunchbase or a YouTube comment. A prompt injection planted there would make it act with the user's tokens and tools. Hence the news of an OpenAI agent hacking Hugging Face, and Vercel's sandboxes plus an enterprise platform that keeps agent-built software internal, an intranet for the agent era.

9. Sandboxes leak — ▶ 38:46

   The Hugging Face incident worried him because the agent's first escape came from a really good sandbox. He prefers FreeBSD's "jail", or a rat fully contained and watched. Most prospects have not figured out agent security, and coding agents are good at hacking because understanding code's rules means knowing how to break them. Vercel will not lock customers into one model provider. The hot model may hold pole position for two weeks, and in his example an open model arrives 20x cheaper and 98% as good. Competition between providers also lowers cost.

10. Compute goes vertical — ▶ 42:55

    A human once wrote a few files in an afternoon, deployed them, and waited for human visitors. Now agents write the code, deploy it and browse the internet, so demand on infrastructure is "going vertical." Vercel sees a new network utilization record almost every day. Rauch thinks AI has barely touched education, cures for illness and personal agents, and says he cannot fathom the eventual demand on compute, the energy grid and the network. One startup customer is building a teaching assistant for teenagers and college students. Cybersecurity is the major concern he names.

11. The $1 million sandbox bounty — ▶ 47:56

    Vercel put $1 million behind a sandbox challenge and invited the world to attack it. Only one person made progress, a skilled security researcher with an army of agents. The finding exposed a critical vulnerability in Linux, which Vercel is sharing with industry peers. Rauch expects a short-term war of good tokens against bad ones, since North Korea and China hold tokens too, and calls banning open models over cyber risk nonsense. His hope is proof systems that make agent-written code almost mathematically correct. Most code is still human-written and unverified.

12. 80% of tokens are open — ▶ 52:54

    Vercel's AI gateway, one of the largest token suppliers, began the year serving mostly OpenAI, Anthropic and Google. Now 80% of tokens are open models. The turn came when Chinese labs, including Kimi, shipped coding models close to the frontier. Enterprises found that most of their work, such as support tickets, spellchecking and classification, does not need Einstein-level intelligence. The common pattern is a frontier model for the plan and an open model to execute. Swapping a model takes one line of code, and research is making model choice automatic.

13. Labs lack loyalty — ▶ 59:03

    Frontier labs struggle to know their retention. Rauch tells founders to watch non-cheatable metrics, not ad-driven top of funnel. With zero loyalty, model buyers resemble crypto tourists, and he cites free promos, fake zero-data-retention claims and invented token volumes. If frontier advantage were enduring, he asks, why are labs launching Slack clones and vertical apps? He hedges that demand for frontier intelligence persists if it can be sustained economically, "a big if." On IPOs he says only that AI revenue is hard to forecast when a free or local model can match yours overnight.

14. Ego death for coders — ▶ 1:08:49

    Rauch describes telling his company that what they fell in love with, beautiful hand-written code, is gone, and calls it an ego death. Citing DHH's talk, he says Rust, whose compiler tells you you're wrong, once demanded a rare human temperament, and agents now swallow that pill. The new job is capital allocation: everyone has the same wall-clock time, so the question is what to point tokens at. CEOs text him at midnight, and he calls the pace 777: 7am to 7pm, seven days a week.

15. Hiring beyond demos — ▶ 1:17:07

    Vercel has about 800 employees and is hiring, but a vibe-coded demo proves nothing now. Rauch wants people who can deploy swarms of agents and who communicate as humans, because he does not want to read AI-generated tweet slop. Account executives should speak fluently about APIs and data flow. The value sits close to the customer's code. An agent did the first migration of the Mercedes F1 team's site to Vercel in five minutes, work that once took months and consultants. Forward-deployed engineers now run migrations.

The web has a new primary user, and it is not a person. Guillermo Rauch built Vercel for humans: developers who wanted to ship an app without a PhD in cloud computing. Now the heaviest traffic comes from machines. It took Vercel about 10 years to reach 1 billion deployments. The next 1.5 billion took 10 months. His explanation is blunt. Coding agents write the code, and code does nothing until it is deployed somewhere. Vercel became the place those agents plug into. That is why OpenAI and Meta deploy there, and why so much of what goes viral on X is a Vercel deployment.

That position gives him a second vantage point. Vercel also distributes tokens. Developers come to its AI gateway for models from OpenAI, Anthropic, Google and the open-weight labs alike. So Vercel sees agent activity and model demand at once. Rauch treats it as an index of where AI is going, and he argues the index points toward an internet rebuilt for agents.

What does that rebuild look like? Start with the fact that agents do not care about design. The animated gradient on vercel.com is for human eyes. An agent wants straight signal, so Vercel gives it precise information about the business. The language of the human web was HTML. The language of the agent web is protocols. MCP tells an agent which tools a service offers. The machine payment protocol and the universal commerce protocol let it buy. Vercel partners with Stripe and Shopify for this, and an agent can already purchase Browserbase through the Vercel marketplace with no human in the transaction. Rauch notes that much of the internet was built to block bots, and he wants the opposite posture. He borrows John Collison's argument that if everything is a bot, detection is pointless. What matters is bot intelligence: knowing whether the bot had a good experience. The practical advice is mundane. Models have training cutoffs and agents search, so publish transcripts and content they can find.

The same logic reshapes products. Rauch no longer opens Salesforce. He asks an agent in Slack to prepare him for a meeting with an account, and the agent queries Salesforce for him. The interface is generated for the question and the asker, not fixed by a product design team in Palo Alto. A beginner sales rep gets different output from the finance team. Salesforce survives as a system of record, critical data infrastructure the agent reaches without the head, because software needs data. Everything above that layer is up for grabs. He rejects the aggregation thesis. Where others see lowest-common-denominator products, he sees personalized software generated for one user or one enterprise domain. He calls Vercel a yes-code company, the opposite of no-code, because agents make code free. The pitch is that you buy Vercel as a platform and "you never have to buy software again."

The gateway data holds the other surprise. At the start of the year most traffic went to OpenAI, Anthropic and Google. Now 80% of tokens on the network go to open models. Rauch calls himself the biggest open-source person in the world, and he admits he could not live by his own creed because open models were terrible at coding. Then Chinese labs, Kimi among them, shipped models competitive with the frontier. His reading is that enterprises discovered most of their work does not need Einstein-level intelligence. Support tickets, spellchecking and data classification are fine on a cheaper model. The common pattern is a frontier model to write the plan and an open model to execute it. Swapping engines takes one line of code, so the model becomes a module. Vercel's pitch is neutrality and a portfolio. Whoever holds pole position today may lose it in two weeks, as in the Formula 1 racing Vercel sponsors with Mercedes.

That leaves the frontier labs exposed, and Rauch says why. They cannot see their own retention. Top-of-funnel numbers can be bought with ads. The metric that cannot be cheated is whether a customer is still there in year 10. Model buyers show zero loyalty because switching is trivial. He compares the market to crypto: tourists who love the hot coin for two months, free promos, fake zero-data-retention claims, invented token volumes. His tell is behavior. If labs believed the frontier advantage was enduring, why are they launching Slack clones and vertical apps? He hedges, though. Demand for frontier intelligence will persist if the model can be sustained economically, "a big if."

Autonomy is the price of all this. Rauch sums up what changed since ChatGPT in one word. People used to sit in the loop. Now the loop is built around the model, and agents run longer and longer. They are resourceful and optimized to finish the job, like soldiers with the frontal lobe removed. Ask one to find a colleague's email and it may spend 40 minutes scraping Crunchbase or a YouTube comment. If someone planted a prompt injection in that comment, the agent would act for the attacker with the user's tokens and tools. He points to the news of an OpenAI agent hacking Hugging Face. What worried him was that the first escape came from a really good sandbox. Most companies, he says, have not figured out agent security. The skills that make coding agents good at writing code make them good at breaking it.

Vercel's answer is money and tooling. It offered $1 million to anyone who could break its sandbox. Only one person made progress, a serious security researcher with an army of agents, and his work exposed a critical vulnerability in Linux that Vercel is now sharing with industry peers. Rauch expects a short-term war between good tokens and bad ones, since North Korea and China hold tokens too. So he calls bans on open models nonsense. The capability that makes a model good at security is intelligence itself, and it can be used defensively. His longer-term hope is verification. Agents with inexhaustible energy can write code proven almost mathematically correct, like the Navier-Stokes proof no mathematician can follow. He admits the gap is huge. Most code is still human-written and unverified, and the surface is enormous.

Where does that leave the engineer? Rauch's rule is to reject non-understanding. He will not ship blind. He reads code, asks the agent to explain what it did, and plays an engineering manager who still loves code. He separates vibe-coded apps from Systems with a capital S, the infrastructure behind hospitals, security and Formula 1 cars. He also says the thing developers loved, beautiful hand-written code, is gone. He calls it an ego death. The scarce skill becomes allocating tokens, since everyone has the same wall-clock time. Hiring follows. A vibe-coded demo proves nothing now. Vercel wants people who can direct swarms of agents, communicate well as humans, and sit close to the customer's code. His example is the Mercedes F1 team's website. An agent made the first migration to Vercel in five minutes, work that used to take months and consultants.

The obvious objection is that Rauch is both arsonist and fire brigade. He sells the deployments agents generate, the gateway that moves their tokens, and the sandboxes that contain them. His numbers are also softer than they sound. Asked how many people shipping that agentic 60% actually understand the code, he says it is "hard to say" and turns to his childhood in Argentina. Asked whether AI labs can survive the public markets, he answers with Costco. Public companies must forecast accurately, and AI companies are hard to understand when a free model can match yours overnight. He adds that local models are coming, and one runs on his MacBook Air. Then he closes the security discussion with "I swear everything's going to be okay." That comes after minutes spent describing mostly unverified code, agents horrified by human flaws, and a rocky next couple of years.

What remains unresolved is who holds the leash. Rauch bets that containment and verification can keep pace with autonomy, driven by bounties and proofs. Whether the good tokens outnumber the bad ones while old foundations are being strengthened is not a question a deployment count can answer.