This is a recap built from Vercel's published event materials, its changelog and the session program. Figures marked "according to Vercel" are the company's own numbers and I haven't verified them independently.

On October 15, 2026, Vercel held Ship SF at the Palace of Fine Arts in San Francisco. According to Vercel, more than 2,700 developers, founders and technical leaders attended. It was the last stop in a five-city Ship series that also visited London, Berlin, New York and Sydney, and Vercel says the whole series drew over 10,000 people.

Agents were the subject of nearly every session. But the most useful lesson wasn't about chatbots. It was how much ordinary engineering sits underneath an agent people actually rely on: storage, background workers, identity, budgets and security. Here is what was announced, who spoke, and what it adds up to.

The big question: will AI replace developers?

Guillermo Rauch, Vercel's founder and CEO, opened by taking on the fear that keeps coming up. He answered with data from Vercel's own platform. According to Vercel, over the last 12 months:

• The number of developers on the platform tripled.
• The number of new projects they ship doubled.
• Deployments triggered by agents grew 28 times over. His argument was that AI hasn't pushed developers out. More people are building, they're shipping more, and agents are doing a growing share of the deploying. He explained Vercel's focus on developers by saying they are what separates "a world full of slop and a world full of software worth using."

He also sketched where he thinks software is heading: interfaces generated around whatever task you're doing, agents that find and fix bugs on their own, and products that improve from what they learn in production. In his view, real-world production data is what turns software into a feedback loop.

What Vercel means by "agentic infrastructure"

Vercel describes its platform as serving agents in three ways:

1. A place to build and run your own apps and agents, with the production tools included.
2. A place where coding agents deploy. Vercel says it exposes more than 200 CLI commands so an agent can operate the platform itself.
3. A platform that is itself run by agents. Instead of only alerting you to a problem, Vercel's agents investigate it and open a pull request. Underneath sits what the company calls its Agent Stack:

Backends and background work: Daemons

Vercel's CTO, Malte Ubl, argued that Vercel is no longer only a home for frontends. In his keynote he showed APIs, backend services, containers, MCP servers and background workers all running on the platform. He also announced that Python support is now generally available, covering FastAPI, Flask and Django with standard CPython, PyPI packages and uv. Combined with Vercel Services, a frontend and its backend can deploy together and be previewed as one app.

The announcement that stood out most was Vercel Daemons, now in public beta. A daemon is a process that simply stays alive. You deploy your code or container image, say how many instances you want, and Vercel keeps that number running, replacing any that crash. Vercel's changelog lists the kinds of work this suits:

• Queue consumers built on tools like Celery, BullMQ, RabbitMQ or Kafka
• Temporal or Cadence workers that long-poll for tasks
• Self-hosted CI runners that pull jobs instead of accepting inbound connections They run as containers on Fluid Compute, inside the same project as your app, and they don't need an HTTP server.

This matters because serverless was always an awkward fit for work that never stops, and agents create a lot of that kind of work. Daemons fill that gap.

Persistent agent work: Drives, eve and Connect

Tom Occhino, Vercel's chief product officer, covered the Agent Stack and its newest parts.

Vercel Drives (generally available for Sandbox) address a real limitation. A sandbox is temporary, but some agent jobs aren't. Research and large refactors can take days. A Drive keeps the agent's code, data and notes in place between runs, so you can swap the runtime or compute underneath without starting over. Imagine an agent halfway through a repository-wide refactor: its progress file and edits are still there even after the sandbox that made them is gone.

eve is Vercel's open-source agent framework. The simplest comparison is that eve is to agents what Next.js is to React: a framework with sensible defaults so you don't assemble everything yourself. New this time are memory that persists across sessions, installable extensions for tools and skills, and the ability to create and deploy an agent from the Vercel dashboard.

Vercel Connect is now generally available, with more than 1,000 connectors. It gives agents scoped access to outside tools and data. New in public beta is MCP Events, which lets an agent react when an MCP server sends a webhook, such as a new message, a newly opened issue or feedback on a document. Connect manages the subscriptions and renewals, and delivers events securely.

AI Gateway gained spending controls. You can set budgets per API key, project, person or team, restrict inference to the United States, and turn on Zero Data Retention so model providers don't keep your prompts or responses. For anyone who has watched an agent loop run up a bill, budgets are one of the most practical announcements of the day.

Vercel also introduced Token Pass in early access for Pro teams: $80 of GLM 5.3 and Kimi K3 credit for $30 a month, with more models planned.

Vercel runs on its own agents

Vercel presented itself as its own customer. Matt Lewis from Vercel and Mala Neti from Slack explained how Vercel built "v," its internal agent. It lives in Slack, uses Vercel's own data, reaches tools like Snowflake, GitHub and Linear through Connect, and requires company sign-in through Passport.

The number most people repeated was about support. Vercel reports that its support agent, Vertex, handles about 91% of support queries. In a session at the London event, Vercel's team said the agent automates 91% of tickets and saves roughly 5,000 engineer-hours a month. These are company-reported figures, but they show how deeply Vercel is putting agents into everyday operations.

Enterprise: governing what everyone builds

Jeanne Grosser, Vercel's COO, described a problem every large company is about to face. When engineers, marketers, analysts and salespeople can all build software, someone has to decide who can ship, who can open an app, what systems an agent may touch, and how much it can spend.

Two announcements addressed that:

• Bring Your Own Cloud on AWS is generally available. Your application functions run inside your own AWS account, within your own security and network boundaries.
• Enterprise-Managed Authorization for Vercel Connect is in public beta. Once a supported service and your identity provider are set up, agents can act for employees without each person approving every connection one by one. Passport, which places internal apps and agents behind a company's identity provider, rounds out the picture. A session from Mercor showed what this looks like in practice. Aaron Brown and Arty Vohmincevs of Mercor explained how hundreds of employees build internal tools without losing control of access or spending: apps inherit a managed identity layer, and AI Gateway gives the team visibility into model usage and budgets.

Voices from other companies

The program mixed Vercel's own launches with talks from other teams:

• Slack and Notion. Beyond the Slack session on "v," Sarah Sachs of Notion joined Malte Ubl to discuss how AI is changing the software teams build and what context agents need to do useful work.
• Meta. Prashant Ratanchandani of Meta Superintelligence Labs talked about building products, including Muse, on Vercel. The day before Ship, Vercel and Meta co-hosted a hackathon called Fast Forward, supported by Shopify and Neon, with tracks for long-running agents, multimodal apps and VR.
• Anthropic. Katelyn Lesse covered what comes after an agent is live: keeping it fast and improving it across thousands of production sessions with Claude Managed Agents on Vercel.
• SAP. Philip Miseldine and Sophia Levens showed generative interfaces built with v0, eve and json-render. Their emphasis was consistent design and meeting government requirements, a sign that generated UI can work in regulated settings.
• Supabase. Gurjeet Bhangu of Supabase described the company as "the Vercel backend partner." Through the Vercel Marketplace, an agent app can be connected to Postgres, auth and storage in a few steps.
• Adapta. Max Peters described the router behind Brazil's largest generative AI platform. Built on AI Gateway, it offers 168 models through one integration, picks a model in under 200 milliseconds, and handles more than three million requests a day with automatic failover and per-user budgets.
• Ora. Assaf Elovic showed how his team tests whether websites work for agents by sending real agents to sign up, integrate and pay, then tracing each step.
• Others. AWS covered production patterns for agents, ClickHouse spoke about how open source earns developer trust, and a "Fast AI" panel with the CEOs of TypeSafe AI and Cerebras discussed speed, reliability and cost. ## Shopify and "The Third Customer"

The day ended with Guillermo Rauch in conversation with Shopify CEO Tobi Lütke, under the title "The Third Customer." The idea is building for a world where agents research, compare and buy on people's behalf.

A few of Tobi's points:

• Project-based shopping. Agents can help people follow through on things they never had time for, like planning a Halloween setup or a 3D-printing project with their kids and getting everything delivered before the weekend.
• Good storefronts survive. Distinctive, well-made shopping experiences can still grow. The ones in trouble are the undifferentiated ones.
• Verification engineering. Shopify's engineers increasingly build test harnesses and acceptance checks that make AI-written changes verifiable. One example is a rewrite of the storefront renderer from Ruby to Rust, where outputs are compared byte for byte before switching over. Deleting code is celebrated too.
• A tip for builders. Explore Shopify's catalog APIs, which open up products from Shopify merchants, so you can build discovery and curation tools without first recruiting both sides of a marketplace.

What I take from this

1. Agents are turning into ordinary software. Like any production system, they need hosting, storage, queues, identity and cost controls. Drives, Daemons, Connect and Passport are Vercel's answer to those needs.
2. Security and cost are product features now. Short-lived credentials, enterprise authorization, bring-your-own-cloud, spend budgets and zero data retention all appeared in one program.
3. Developers aren't disappearing. Vercel's growth numbers point to the work shifting toward directing, verifying and shipping. Tobi Lütke's "verification engineer" label fits that shift.
4. Your next customer may be an agent. The commerce and agent-testing sessions both pointed the same way: design products so agents can use them, with clear APIs, scoped access and good documentation. If you build on any platform, the practical question to take away is this: what happens to your agent's work when the process running it stops? The announcements at Ship SF were mostly different answers to that question.

Sources and further reading

• Vercel Ship 2026 recap
• Vercel Daemons public beta
• Vercel Drives generally available
• MCP Events support in Vercel Connect
• Bring Your Own Cloud on AWS generally available
• Enterprise-Managed Authorization for Vercel Connect

I'm an independent writer and this post is not affiliated with or sponsored by Vercel.