Vercel's Agentic Infrastructure

LLMs and coding agents are shifting the final actor in software development from humans to machines. Vercel revealed that over the past 3 months, weekly deployments have doubled, and over 30% of all deployments now originate from coding agents. This is a 1000% increase from 6 months ago, with Claude Code leading at 75%, followed by Lovable and v0 at 6%, and Cursor at 1.5%.

The new infrastructure Vercel describes is Agentic Infrastructure. This shift breaks down into three stages:

• Infrastructure that coding agents can deploy to
• Infrastructure for building and running agents
• A structure where infrastructure itself behaves like an agent

The core problem is operational friction. Even when a coding agent writes a feature, it needs a URL to run, test, and verify it. If that path involves manual Terraform state management or cloud console clicks, the autonomous loop breaks. This is why immutable deployments, preview URLs for every commit, and instant rollbacks are no longer conveniences but prerequisites for machine-driven development.

Through CLI, API, MCP servers, and git integration, Vercel provides a surface where agents can write code themselves, open PRs, verify with preview URLs, and deploy all the way to production. Deployment is no longer a process manually stitched together by humans, but a deterministic execution path that agents can directly operate.

Agent workloads differ in nature from traditional serverless. They require long-lived execution, multi-step orchestration, model routing, cost controls, sandboxed code execution, and abuse resistance all together. Running this stack independently quickly leads to wasted inference costs, provider outages, and untrusted code risks piling up.

To address this, Vercel bundles several AI primitives into a single platform.

• AI SDK: a unified approach to AI development across frameworks and providers
• AI SDK 6: adds agent abstraction, allowing a single agent to be reused across multiple interfaces and workflows
• Chat SDK: extends agents across multiple chat apps and platforms
• AI Gateway: connects hundreds of models through a single endpoint, providing budgets, monitoring, routing, retries, and fallbacks
• Fluid compute: an execution environment tailored to the unique latency, concurrency, and idle waiting patterns of AI workloads
• Workflows and Queues: pause, resume, retry, maintain state, and separate background work
• Sandbox: an isolated execution environment for untrusted code
• Observability: tracks agent behavior and points of failure

What sets this configuration apart isn't simply a bundle of features but shared context. By viewing code, model calls, and runtime behavior together within one system, infrastructure moves beyond observing the production environment toward interpreting and responding to it on its own. In traditional infrastructure, code went in, logs came out, and humans read those logs to fix things — but a unified platform holds all the context in between.

When a latency spike or model provider outage occurs, the platform investigates the anomaly, checks observability data and logs, and reviews the source code to perform root-cause analysis and fix review. Today, human approval still remains in the loop, but over time the platform is expected to take on more of the operational burden. If the cloud detached humans from the system, the next stage is infrastructure built on the premise that software writes, deploys, and fixes itself.