Vercel Says Coding Agents Now Drive 30% of Deployments

Coding agents now start more than 30% of all deployments on Vercel — up 1000% from six months ago, by the company's own count. That pace matters to any engineering leader currently planning tooling budgets or headcount around how software actually gets shipped.

What changed

In a blog post titled "Agentic Infrastructure," Vercel says weekly deployments on its platform have doubled in just three months, and coding agents are driving nearly all of that growth. The company reports that over 30% of deployments are now initiated by coding agents, up 1000% from six months ago.

Vercel also broke down which agents are doing the work. According to the company, here is how agent-initiated deployments split by tool:

Coding agent	Share of agent-initiated deployments
Claude Code	75%
Lovable + v0	6%
Cursor	1.5%

Vercel did not disclose what accounts for the remaining share or its methodology for attributing a deployment to a specific tool.

The company also claims that projects deployed by coding agents are 20 times more likely to call AI inference providers than projects deployed by humans — evidence, Vercel argues, that agent-shipped software looks structurally different from human-shipped software, not just faster to produce.

Why Vercel is betting on "agentic infrastructure"

Vercel frames this shift as the reason it needs three layers of what it calls agentic infrastructure: infrastructure for coding agents to deploy to, infrastructure for building and running agents, and infrastructure that is itself agentic. In the company's words, the bottleneck for agentic engineering is operational friction — when an agent writes a feature, it still needs somewhere to run, test, and get a URL for the result.

To close that gap, Vercel points to its existing CLI, API, MCP servers, and git integration, which it says let agents open a pull request, generate a preview URL, verify the output, and ship to production without a human clicking anything. It also cites AI SDK 6, which adds an agent abstraction so a team can define an agent once and reuse it across products, plus AI Gateway, a single endpoint for routing requests across models with budgets, monitoring, and fallbacks built in. Vercel also points to Fluid compute, which it designed for the unusual shape of AI workloads, where latency, concurrency, and idle waiting time all have to be managed simultaneously.

Read as a product pitch rather than a research paper, the through-line is simple: fewer manual steps between an agent's code and a live URL.

A concrete example: the new Vercel plugin for Claude Code and Cursor

A companion release gives a clearer sense of what agentic infrastructure looks like day to day. Vercel shipped a plugin that, in the company's description, lets Claude Code and Cursor further understand Vercel projects by watching file edits and terminal commands in real time and injecting relevant platform knowledge into the agent's context as it works.

Per Vercel's release notes, the plugin bundles access to more than 47 skills covering Next.js, the AI SDK, Turborepo, Vercel Functions, and routing middleware, drawn from what the company calls a relational knowledge graph of the platform. It also ships three specialist sub-agents — an AI Architect, a Deployment Expert, and a Performance Optimizer — plus five slash commands: /bootstrap, /deploy, /env, /status, and /marketplace. Vercel says the plugin compiles pattern matchers at build time and runs a priority-ranked injection pipeline across seven lifecycle hooks, rather than relying on standard retrieval.

What this means — and what it doesn't

Read generously, these numbers are a leading indicator. Engineering leaders sizing headcount or evaluating coding-agent budgets now have a large, self-reported data point suggesting the shift from manual to agent-initiated deployment is happening faster than most roadmaps assume.

Read skeptically, this is Vercel marketing its own platform using its own metrics, published with no independent audit and no disclosed methodology for how a deployment gets attributed to Claude Code versus a human simply using Claude Code inside an IDE. An agent "initiating" a deployment is also not the same as an agent shipping unsupervised: Vercel itself says that when its systems detect anomalies like a latency spike, they investigate, read logs, and propose fixes today — but a human still approves the change before it goes live.

That distinction matters for anyone reading "30% of deployments" as proof that agents are replacing engineers. Vercel's own framing is closer to agents removing operational friction from work engineers still direct, not engineers stepping out of the loop entirely. For teams that already run on Vercel, the practical question is less about the topline percentage and more about whether their own deployment logs would show a comparable shift if they looked.

What to watch

• Whether Vercel publishes the same deployment breakdown again next quarter, and whether Claude Code's 75% share holds, grows, or gets contested as Cursor and other agents ship competing integrations.
• Whether the human-approval step Vercel describes for incident response narrows over time, as the company suggests it will, or remains a permanent guardrail.
• Whether other cloud and hosting platforms publish comparable agent-initiated deployment numbers, which would make Vercel's figures easier to verify against an industry baseline instead of standing alone.