One trillion tokens a day. That's the number Guillermo Rauch drops casually, the way other CEOs cite quarterly revenue. Six million deployments daily. Half of them triggered not by developers but by coding agents. Vercel didn't set out to become the control plane for the agent economy. It built a deployment platform. The agents showed up anyway.

Rauch knows the timeline better than most. Last year was the prototype phase — "unleash the agents, everyone can build." Vercel did that internally, hundreds of agents sprouting organically. Then came the hangover: agents in production behave differently than agents in demos. They leak data. They hallucinate tool calls. They resist audit. The sky-is-the-limit rhetoric hit the ceiling of compliance, security, and operational reality.

The two killer apps

Rauch distills the noise into two use cases that actually matter. First: coding agents. Obvious, yes, but the scale is not. When agents write code at this velocity, you don't need a better IDE. You need a place to put the software. That's Vercel's existing business, supercharged.

Second: internal corporate agents. The sales rep growing existing accounts. The procurement officer matching invoices to contracts. The bottleneck for these people was never creativity — it was access. Access to the right data, the right systems, the right context, without violating policy or exposing crown jewels. That's where the fight gets interesting.

The cage and the contract

Vercel's answer arrives in two pieces. Eve lets you write agent instructions and skills in natural language — a contract, essentially, between human intent and machine execution. Sandbox puts the agent in a cage. It can reason, it can call tools, it can express intelligence. But data egress? Policy-enforced. Access controls? Auditable. Every tool call leaves a trail.

The Airbus anecdote Rauch tells is the point of the spear. Decades of aerospace C++ code. One developer installs the wrong IDE plugin — Cursor, Devin, pick your poison — and the entire codebase ships off to someone else's training run. That's not a hypothetical. That's a board-level risk. Sandbox exists because the model providers won't solve it. Their incentive is consumption. Vercel's incentive is trust.

Models are commoditizing. The agent layer isn't.

This is the fight Rauch is describing without using the word "fight." The major labs — OpenAI, Anthropic, Google — want to own the full stack. Model, tools, memory, deployment, governance. Vercel wants to split that stack. Models become interchangeable commodities. The agent execution environment — the sandbox, the policy engine, the audit log, the deployment target — becomes the moat.

You see it in the token volume. One trillion daily tokens flowing through Vercel's AI gateway. That's not inference; that's orchestration. Routing. Guardrails. Billing. The gateway sits between the agent and the model, and that position is worth more than the model itself. Ask any cloud provider who won the container wars. It wasn't the container runtime. It was the control plane.

The platform paradox

Vercel competes with the labs now. Not on model quality — on agent infrastructure. The labs build agents (Operator, Computer Use, Devin). Vercel builds the platform those agents need to run in production without getting sued or fired. The labs will inevitably build their own deployment, their own sandboxes, their own governance. They have to. Vertical integration is the only way to capture the full margin.

Rauch knows this. He's watched platforms absorb their ecosystems before. The question is whether Vercel's head start — six million deployments, a developer base that already treats Vercel as default infrastructure — creates enough gravity to hold the line. The agents are already here. They're deploying. They're generating tokens. They're touching production data. Someone has to provide the cage.

For now, that someone is Vercel. The fight to split models from agents isn't theoretical. It's measured in trillions of tokens a day. And it's just getting started.