AI Agents Now Trigger a Third of All Deployments on Vercel

The number of Vercel deployments triggered by AI agents has doubled in three months. Agents now launch 30% of all projects, up from closer to 3% a year ago. It's a huge jump, and it has exposed an old problem: the infrastructure was designed for humans. It assumed someone would configure servers, click the deploy button, and read the logs. Now the actor is a machine, and that calls for a completely different approach.

Machines code faster

Projects deployed by agents are 20 times more likely to call AI services (such as OpenAI, Anthropic, or Claude) than projects written by humans. In other words, AI agents write code that uses AI, and they often write code for other AI agents. This software runs at full speed: it tests itself and ships to production without waiting for approval.

Claude Code accounts for 75% of these automated deployments on Vercel. It has moved beyond code editing to become a full-fledged infrastructure player. Lovable and v0 follow (6%), then Cursor (1.5%). The key point is that these numbers have grown 1000% in six months. What used to be the exception is becoming the rule.

Infrastructure has to change

When a human writes and deploys code, they need control and visibility. When an agent does it, it needs speed. If clicks in a cloud web console or manual Terraform management stand between the code and the production system, the autonomous loop breaks. Agents need clean deployment APIs, preview URLs to check results, and instant rollbacks.

That changes the requirements radically. Immutable deployments, instant rollbacks, and preview URLs for every commit are no longer developer conveniences; they are an absolute necessity for machine-driven development. Without them, an agent simply can't verify what it wrote.

The second layer of complexity: agents themselves become long-lived workloads. They need multi-step orchestration, routing between models, and cost control. Every API call, every request through AI Gateway, every execution in Sandbox costs money. The infrastructure has to pause, resume, retry, persist state, and log everything that happened.

Vercel's three-layer solution

Vercel has put forward an integrated toolstack:

• AI SDK — a single foundation for building AI applications. Version 6 adds an agent abstraction, so you define an agent once and use it everywhere
• Chat SDK — access to dozens of chat platforms from a single codebase
• AI Gateway — a single entry point to hundreds of models, with budgets, monitoring, fallbacks, and retry logic
• Fluid Compute — optimized for the unusual phases of AI workloads (latency, concurrent requests, idle time)
• Workflows and Queues — pause, resume, retry, and state persistence
• Sandbox — an isolated environment for untrusted code
• Observability — monitoring of agent costs and behavior in production

What it means

For fifty years, infrastructure assumed a human would configure it, click the deploy button, and read the logs. Each generation of software demanded a new generation of infrastructure. The cloud turned infrastructure into an API. Then infrastructure started being derived from the application itself.

Now a third shift is underway: infrastructure is adapting to the fact that its end user is not a human but an autonomous machine. And it isn't happening slowly. Everything changed in three months.