Vercel Ship 2026: Agents Now Drive Half of Deployments

Vercel Ship 2026 made the agentic infrastructure shift impossible to ignore: the company said that the share of deployments triggered by coding agents has gone from under 3% six months ago to more than half today — what Vercel framed as the fastest infrastructure shift it has recorded. The numbers are vendor-stated, but the direction of travel is the real headline.

Held in London on June 17, 2026 in front of more than 2,500 attendees, the keynote was Vercel’s first Ship conference outside the United States. CEO Guillermo Rauch opened with a line that set the tone for the day, and the announcements that followed were less about new framework sugar than about a single hard question: once agents are writing and shipping most of your code, how do you run them in production securely, at enterprise scale, without handing a shared bot token the keys to everything?

This guide does three things. It plots the deployment-share curve across the only three public data points that exist — and explains why the slope, not the headline, is the story. It maps the new enterprise stack — Connect, Passport, Services, Agent, and Bring Your Own Cloud — against the security problems each one solves. And it translates all of it into what marketing and engineering teams should actually do now, with every figure attributed to its source.

1. 01
   Agents went from a rounding error to the majority.According to Vercel, agent-triggered deployments rose from under 3% six months ago to more than 50% as of June 17, 2026 — a single-source internal metric, corroborated by Vercel's own press release rather than an independent audit.
2. 02
   AI Gateway token volume grew roughly 10x.Vercel says AI Gateway processing climbed from about 2 trillion to roughly 20 trillion tokens per month over the same six-month window — a second vendor-stated signal that agentic traffic, not just deploys, is compounding.
3. 03
   The bet is OIDC-first identity, not shared tokens.Connect, Passport, Enterprise Managed Users, and Vercel Agent all assume agents need their own verifiable identity. Connect swaps long-lived provider tokens for short-lived, task-scoped credentials issued at runtime.
4. 04
   eve is the open-source framework underneath.Vercel launched eve, an open-source TypeScript agent framework where an agent is a directory of files — markdown instructions plus TypeScript tools — built on the Workflow SDK, Vercel Sandbox, and the AI SDK.
5. 05
   Vertex is the proof point CTOs will quote.Vercel says its internal support agent now automates 91% of support tickets and saves 5,000 engineer-hours per month. It is a vendor-stated figure with no third-party audit, but it is the kind of number that anchors enterprise buying cases.

The single most-quoted number from the keynote is also the one to handle most carefully. According to Vercel, six months ago fewer than 3% of deployments on the platform were triggered by coding agents; as of June 17, 2026, that share is now more than 50%. This is a Vercel-stated metric drawn from the company’s own internal data, corroborated by its press release rather than by any independent measurement — so treat it as a credible directional signal from the company best positioned to see the traffic, not as an audited industry statistic.

That caveat matters, but it does not blunt the point. Even a vendor’s own deployment logs are hard to argue with on direction: the platform that hosts a large share of the modern web is telling you that machines, not people, now click deploy most of the time. For a marketing or product team, the practical takeaway is that “ship velocity” is decoupling from headcount. The work that used to gate a release — wiring, glue code, the unglamorous plumbing between a design and a live URL — is increasingly done by an agent, and the human role is shifting toward review, judgment, and direction.

Rauch’s framing in the accompanying press release is the architecture thesis behind the stat: each new generation of software, he argued, needs a new generation of infrastructure — and for the agent era, that infrastructure is where coding agents deploy software, where teams build and deploy their own full-stack apps and agents, and increasingly a place that can autonomously monitor software once it is live. The deployment number is the evidence; the new product stack is the response.

The reason the headline percentage is less interesting than the curve is that there are actually three public data points, not one. The keynote gave the endpoints — under 3% six months ago, over 50% now — but an earlier, undated Vercel post on agentic infrastructure (which the data places at roughly April 2026) recorded an intermediate reading: more than 30% of deployments initiated by coding agents, up around 1000% from six months prior. Lined up, those three readings describe an inflection, not a step.

Do the arithmetic on the two endpoints and the picture sharpens. The token figure is the cleanest: roughly 20 trillion divided by roughly 2 trillion is a 10x increase in six months, which is exactly how Vercel describes it. The deployment share is messier to multiply because both endpoints are bounded (“under 3%” and “over 50%”), but even taking the conservative read, going from below 3% to above 50% is more than a sixteenfold rise in share over the same window. Two independent signals — deploys and tokens — moving an order of magnitude in the same direction is a stronger argument than either number alone.

One honest note on the middle data point: the agentic-infrastructure post carries no visible publication date, so we anchor it at “circa April 2026” based purely on where its 30%-plus reading falls between the December and June endpoints. We have not assigned it a precise date, and the 1000%-growth and related figures it contains should be read as approximate rather than timestamped.

The interesting structural argument is not that agents crossed 50% — it is that the curve has a self-reinforcing mechanism, so there is no obvious reason it stabilizes there. Several of Vercel’s data points, taken together, describe a recursive loop in which agents build software that uses agents, which then generates more agentic traffic.

Start with the inference multiplier. In the agentic-infrastructure post (circa April 2026), Vercel reported that projects deployed by coding agents are roughly 20 times more likely to call an AI inference provider than projects deployed by humans. Agents do not just ship code; they ship code that itself calls models. Layer on the internal evidence Vercel shared at Ship: its internal data agent, d0, now gets about 45% of its questions from other agents rather than from people, and its support agent, Vertex, automates a vendor-stated 91% of tickets. The platform is already running a loop where agents query, build, deploy, and support agents.

Our read: this is the structural case for why agent share is unlikely to plateau at half. When the entities doing the deploying are themselves 20x more inference-hungry, and a growing slice of platform queries already originate from other agents, each turn of the loop adds more agentic demand than it removes. We would not bet on the exact slope — vendor metrics compound their own optimism — but the mechanism is real, and any team modeling its own tooling and hosting costs over the next year should assume agentic traffic keeps growing faster than human traffic, not that it levels off.

If agents are now the majority of who deploys, the unsolved problem becomes governance: an autonomous process needs credentials, network access, and the ability to act — and handing it a long-lived shared token is exactly the wrong model. The through-line of every enterprise announcement at Ship 2026 is a single thesis: agents need their own verifiable identity, issued and scoped through the same OIDC machinery enterprises already trust for humans, not a bot account with standing access.

The table below maps the enterprise security surface — the capabilities a team would otherwise have to roll itself — against the Ship 2026 product that addresses each, with the announced availability status and any stated pricing. Note carefully where pricing is and is not public: Connect has published rates, while Passport and the private-beta products are quoted as enterprise pricing by arrangement with Vercel.

Vercel Connect, now in Public Beta, is the clearest expression of the OIDC-first thesis. Instead of storing a long-lived provider token in your app, the app proves its identity via OIDC at runtime and receives a short-lived, task-scoped credential from the provider. At launch it supports Slack, GitHub, Linear, Discord, Notion, Salesforce, Figma, and Snowflake, plus generic OAuth and API-key connectors; Resend, Workday, and Microsoft Teams are listed as coming soon. Connect’s trigger forwarding, in beta, goes a step further — it verifies incoming webhooks from Slack, GitHub, and Linear and re-attests them with the deployment’s OIDC identity before forwarding, so you never store a provider signing secret in the app.

Vercel Passport, also in Public Beta, applies the same logic at the front door. It gates every deployment behind the team’s identity provider by default — Okta, Microsoft Entra, Auth0, or any OIDC-compatible IdP. Pair that with Enterprise Managed Users (Private Beta), which provisions Vercel seats automatically through the customer’s existing directory via SAML SSO and Directory Sync and produces a single audit trail across Vercel and v0, and the shape of the bet is obvious: every actor on the platform — human or agent — is an identity the enterprise already governs.

Underneath the enterprise products sits the developer story. Vercel launched eve, an open-source TypeScript agent framework where an agent is, refreshingly, just a directory of files: instructions in markdown, tools in TypeScript. Durable execution, sandboxed compute, human-in-the-loop approvals, subagents, and evals ship out of the box. eve is built on the Vercel Workflow SDK for durable, automatically checkpointed execution, Vercel Sandbox for isolated microVM execution of agent-generated code, and the Vercel AI SDK for model calls — the same open-source TypeScript agent framework we cover in depth separately, and a credible new entrant alongside the open-source agent frameworks teams are already weighing.

The broader Agent Stack — Vercel’s end-to-end set of building blocks — comprises the AI SDK, AI Gateway, Workflow SDK, Vercel Sandbox, Chat SDK, and Connect. The AI Gateway is worth singling out: it routes across hundreds of models from a single endpoint, fails over automatically when a provider goes down, charges the provider’s own price with no markup, and lets customers bring their own keys. That is the piece processing the roughly 20 trillion tokens a month Vercel cited, and it slots directly into the kind of multi-provider routing strategy we recommend in our modern web development guidance.

The most quotable enterprise evidence at Ship was Vercel’s own dogfooding. The company says Vertex, its internal support agent, now automates 91% of support tickets and saves 5,000 engineer-hours per month — and, notably, that it does not degrade as its context grows. Vercel says it runs more than 100 internal agents and apps on the same Enterprise Apps and Agents platform it announced at the conference. These are vendor-stated figures with no third-party audit, so weigh them as a vendor case study rather than a benchmark — but they are exactly the kind of internal-automation numbers a CTO will reach for when building the case for an agent platform.

Customer voices on the “Agents in Production” panel pointed the same direction. Cursor’s Arthur Viegers argued that autonomy should track risk, and that the better an agent can assess the risk of a change, the more you can let it run unsupervised — citing Shopify and Amplitude, which he said already auto-review and merge a substantial share of low-risk pull requests with no developer time. It is a directionally useful data point precisely because it comes from a third-party practitioner describing other companies’ production results, not from Vercel’s own logs.

The structural lesson under both the Vertex numbers and the Cursor panel is consistent: the unlock is not raw model capability, it is calibrated autonomy. Agents earn more independence as the system gets better at scoring the reversibility and blast radius of each action — which is precisely what the enterprise stack is built to enforce. This is the same governance-first framing we see across large rollouts, from KPMG’s 276,000-person Agent 365 deployment to the broader enterprise AI agent adoption data now landing across the industry.

Most of what Vercel announced is beta or future-dated, so the right posture is not to migrate today — it is to position. The decisions below sort by who you are and what you are optimizing for, with the beta caveats baked in.

For most teams, the highest-leverage move this quarter is the cheapest one: pick a single integration that currently relies on a stored, long-lived provider token and pilot Vercel Connect to replace it with runtime credential exchange. It is low-risk, it is in public beta, and it directly retires the class of secret that agentic workflows make most dangerous. Everything else — Services, Agent, BYOC — is worth tracking on the changelog, not building around yet.

If you are weighing how much of this stack to adopt versus keep portable, that trade-off — lock-in against velocity — is exactly the kind of decision our AI digital transformation engagements are built to work through, mapping the agent platform to the specific governance, residency, and cost constraints your business actually has.

The story isn't that agents crossed 50% — it's that nothing in the curve says they stop there.