Vercel puts its agent infrastructure play on the table at Ship 2026

Vercel used its annual Ship conference on Wednesday to roll out an entire agent platform for enterprise teams. The pieces are Vercel Services, an Agent Stack, the open source eve framework, a managed Vercel Agent, and an enterprise tier. The bet: agent infrastructure should be the same infrastructure that ships human-built software.

The numbers Vercel shared on stage explain why the company is moving so fast. At the start of 2026, fewer than 3% of deployments on Vercel were triggered by coding agents. As of this week, agents are responsible for more than half of all commits. Token volume through the Vercel AI Gateway grew from about 2 million to 20 million over the same period. "Each new generation of software needs a new generation of infrastructure," Vercel founder and Chief Executive Guillermo Rauch said at the conference. "For the agent era, that's Vercel."

What Vercel shipped for agent infrastructure at Ship

The headline launch is Vercel Services, a project-package deployment engine that lets a team ship a full application, including the front end, the back end, and any supporting services, from a single commit, with one URL for the entire app. The company is positioning this as the unit of deployment that coding agents are now expecting to produce. A single Next.js front end plus a Python back end, with routing and authentication across private networks handled automatically, lands in one place rather than as a set of stitched-together repos. Vercel Services enters beta on July 1.

Vercel Connect, the newest addition to the Agent Stack, replaces long-lived API credentials with short-lived tokens that are scoped to the specific people or agents that need them, with a full audit trail. That matters for enterprise security teams because today the typical model is to provision a service account for an agent, hand it a static token, and hope the token does not leak. Connect launches with native integrations for Slack, GitHub, Snowflake, Salesforce, Notion, and Linear, with connectors exposed through OAuth and standard APIs.

eve is the open source piece. Vercel calls it an opinionated framework for building, running, and evaluating agents, and the company is open sourcing the same framework that it uses internally to run its own agents. The bet is the same one Replit made with its agent runtime, and the same one the team behind Databricks Omnigent made with their meta-harness for coding agents: the agent framework itself is becoming a competitive layer, and the team that controls the framework controls how the rest of the stack is used.

Vercel Agent is the managed counterpart. It is an always-on agent that watches traffic, traces, and alerts, then investigates issues and proposes fixes for human approval. Vercel frames it as a "mechanic" that keeps the application and the agent fleet humming. It is available in beta starting today, and the company is pitching it directly to platform teams that have been writing the same homegrown SRE agent on top of OpenAI or Anthropic.

The enterprise pitch: control, identity, and your own cloud

The new enterprise offering, Vercel for Enterprise Apps and Agents, is the part of the announcement aimed at CISOs, platform owners, and IT leadership. The package wraps Vercel's developer tools with identity and access controls and adds the option to deploy inside a customer's own Amazon Web Services account, so an enterprise does not have to give Vercel root over its environment to run agents in production. Enterprise Managed Users give a central team global control over Vercel and v0 seats, and Vercel Passport lets internal apps and agents live behind the company's identity provider by default.

The positioning is the same one Okara, Conductor, and the other Vercel agent customers have been describing in their own case studies this year. Okara runs CMO agents for 120,000 companies on Vercel, and Conductor moved parallel coding agents from the laptop to the cloud using Vercel Sandbox. Both stories hinge on the same idea: agents are running real workloads, and they need an enterprise-grade control plane, not a build-it-yourself stack on top of a model API.

The bigger story is that Vercel is making the same bet on the agent-infrastructure layer that Cloudflare made on the website layer a decade ago. Once you are the place where the work runs, you get to set the rules for how the work gets done. Vercel Connect's short-lived tokens, Vercel Passport's identity-provider integration, and the option to deploy inside a customer's own AWS account are all pieces of that control-plane play. The Agent Stack is the developer-facing surface, and the enterprise package is the buyer-facing surface, but underneath both is the same bet on owning the runtime that agents and their callers depend on. For teams mapping this kind of platform decision against other vendors in the same space, our Best AI Coding Agents in 2026 comparison covers how Vercel's approach stacks up against Cursor, GitHub Copilot, Claude Code, and the rest of the agent framework field.

The catch, and what enterprise teams should actually evaluate

The pitch is strong, but there are three things enterprise teams should pressure-test before they move production agent workloads onto Vercel's new stack.

First, the agent runtime is still a Vercel-managed service. Even with the option to deploy inside a customer's own AWS account, the agent control plane lives in Vercel's own infrastructure, and the agent framework is open source only for the build and evaluation layer, not for the runtime itself. Teams that need a fully self-hosted agent runtime will have to wait, or build their own abstraction layer on top. The full announcement, with the product names and the deployment numbers, is in the original Vercel Ship 2026 coverage from SiliconANGLE.

Second, the credential model is short-lived by default, but only for the integrations Vercel supports out of the box. The launch covers Slack, GitHub, Snowflake, Salesforce, Notion, and Linear, but most enterprise environments have a long tail of internal systems that are not on that list. Until Vercel Connect supports a generic OAuth and service-account layer, enterprise teams will have to either build their own connector or route the long tail through a separate identity provider.

Third, the Vercel Agent is an opinionated agent. It watches traffic, traces, and alerts, and it proposes fixes for human approval. That is a useful default, but it is also a default that platform teams will want to configure. The risk is that Vercel Agent's opinions become a de facto standard for what an SRE agent should look like, and teams that want a different model will spend the first six months unlearning the default.

For enterprise teams that are already on Vercel, the launch is the most natural place to consolidate the agent layer. The combination of Vercel Services, the Agent Stack, Connect, and the enterprise package is the first end-to-end agent platform that covers the full path from a coding agent's commit to a customer's production identity provider. For teams that are not on Vercel, the question is whether the agent platform is the reason to move, or whether the agent platform is a feature that fits on top of an already-decided infrastructure choice.

Either way, the launch sets the bar for what an enterprise agent platform is supposed to look like in 2026, and it is a much higher bar than the model API plus a database that most teams are still running on. The interesting question is no longer whether enterprises will run agents in production. It is whether they will run them on infrastructure that was built for agents from the start, or on infrastructure that was built for websites and is now being retrofitted.