Transcript
00:00:00Hi, everyone. It's so great to be here. Since shipped last year, the world has changed a lot.
00:00:15We're writing less code by hand, but more ideas are coming to life than ever before.
00:00:21In fact, six months ago, less than 3% of our sales deployments were triggered by coding agents or clankers.
00:00:30Since then, the number has grown 17 times. Over half of deployments on Vercel now come from agents.
00:00:37It's wild. But the bigger shift is what those agents are deploying.
00:00:43Since the beginning of the year, agentic workloads on Vercel have doubled.
00:00:49Volume on AI Gateway grew from 2 trillion to 20 trillion tokens per month.
00:00:56And the shift we're seeing is that we're using agents to deploy software that can think.
00:01:02We're shipping agentic sites and apps, but we're using agents to ship agents.
00:01:09It's a really profound change.
00:01:12So I want to go back in time and show you how we got here.
00:01:15Because our new world is agentic, but it all started with websites.
00:01:20Look at that, little G, back in Argentina, building simple sites with HTML, CSS, and JavaScript.
00:01:29I was so excited to put my creations online.
00:01:33And that's what the web is all about, sharing our ideas with the world.
00:01:38In fact, we built that first version of Vercel to make that easy for anyone.
00:01:44It was infrastructure for pages, caching, and global content delivery.
00:01:50Today, we serve the fastest, most beautiful websites in the world for companies like The Met, Supreme, and Warby Parker.
00:02:00But the web is no longer static.
00:02:03It's dynamic.
00:02:05Sites became data-powered applications.
00:02:08So Vercel built infrastructure for servers, APIs, and databases.
00:02:15Now, Zapier, Stripe, and the New York's own today techs run apps at scale on Vercel.
00:02:23For over a decade, we've been building Vercel into a cloud where you can run everything.
00:02:29And in the last year, we've shipped some huge steps towards that vision.
00:02:34We brought on some of the finest Python developers in the world to extend our core infrastructure for back-end frameworks.
00:02:42So now, you can run back-ends like FastAPI, Flask, ExpressJS, and Hono at scale on Vercel.
00:02:53We built out a complete agentic layer.
00:02:56So you can run long-running functions on fluid compute.
00:03:01You can make workflows durable.
00:03:03You can spin up secure sandboxes.
00:03:05And you can host your MCP servers.
00:03:08So you make your app available to any agent.
00:03:13And we brought everything else your app needs right into Vercel, from databases to commerce.
00:03:20So you can manage Amazon Aura, Postgres, and D-SQL, auth providers like Auth0, CMS platforms like Sanity, and checkout from Shopify.
00:03:32All within Vercel.
00:03:35So this means that you can run any architecture you want on our infrastructure.
00:03:41For example, OpenAI, a little AI lab you may have heard of.
00:03:48That's right.
00:03:49They run Next.js front-ends on Vercel, Python back-ends on Vercel, reaching over a gazillion users every day.
00:03:57And so in addition to front-ends, you can host back-end-only services like REST APIs, GraphQL, remember GraphQL?
00:04:08Written in TypeScript, or Python.
00:04:13And so you can deploy workflows and queues that handle long-running, asynchronous jobs, billions of messages every day.
00:04:21You can bring your entire application, and Vercel will run all of it.
00:04:26Which reminds me, people keep asking me, when I'm in San Francisco, they stop me on the streets and they say,
00:04:33what is the Vercel for back-ends?
00:04:35And I tell them, Vercel is the Vercel for back-ends.
00:04:39Of course.
00:04:41And today we'll prove it.
00:04:43The reality is that all of this software is becoming autonomous.
00:04:48It's very, very exciting.
00:04:50Websites and apps used to respond to user input with logic.
00:04:55But now they have agents inside that can understand intent and perform autonomous work.
00:05:02Every new generation of software demands a new generation of infrastructure.
00:05:08So it's natural that agentic software needs agentic infrastructure.
00:05:14And that's exactly what we're building at Vercel.
00:05:17Agentic infra has three parts.
00:05:20Number one, Vercel is where coding agents deploy software.
00:05:25So when you ask Cloud Code or Codex where to deploy, you get Vercel.
00:05:31Because Vercel is built for the way that agents think and work.
00:05:36Second, Vercel is where you build and deploy your own agents.
00:05:41So we give you every tool you need to build and run apps and agents in production securely at scale in one platform.
00:05:51And third, Vercel itself is automated by agents.
00:05:57So Vercel runs your apps in production, we handle traffic, observability, traces, anomalies.
00:06:04And so that data gives our agents the context they need to investigate autonomously and give you PRs, not just alerts.
00:06:13And so for context, teams using Cloud Code deployed over Vercel five times more frequently than teams who don't.
00:06:24That's because we set the standard for developer experience, and now we're doing the same for the experience of the agents that those developers use.
00:06:34Coding agents love deployed over Vercel because we give them everything they need.
00:06:39When an agent needs to verify the work, it needs a live URL you can test.
00:06:45So Vercel gives every deployment a dedicated, secure, preview environment.
00:06:51When you ask your agent to ship an experiment, it needs to roll out that experiment safely.
00:06:58So Vercel gives every experiment a feature flag, the infrastructure to control, and roll out that change with confidence, with instant rollback, and other capabilities.
00:07:08And you don't want your agent wasting time clicking around a dashboard like it's 1955.
00:07:16That's right.
00:07:17An agent is most efficient when every part of the platform is available in its own language.
00:07:24And so Vercel gives the agent a CLI, APIs, MCPs.
00:07:30So this year, as an example, an engineer at Meta opened Cloud Code and asked it to build an internal tool.
00:07:40When it was time to test it, he asked Claude, where do I deploy this internal tool, and push it over Vercel.
00:07:47A week later, everyone on his team was deployed over Vercel.
00:07:51Within a month, Vercel was the go-to platform for all of Meta's superintelligence labs.
00:07:58Meta.ai, Meta's frontier AI product, was born on Vercel.
00:08:05And this happened even though Meta had already spent a bazillion, gazillion, trillion dollars building their own deployment platform, building their own infrastructure.
00:08:13But even the most powerful infrastructure in the world doesn't natively speak the language of agents.
00:08:20Vercel does.
00:08:22And so, agentic infrastructure is where you deploy your websites, apps, and agents.
00:08:30Thank you.
00:08:44G just told you that our ComputeLayup hosts everything on Vercel.
00:08:50I want to tell you about it because for years, running serverless Compute meant living with trade-offs.
00:08:57Serverless was supposed to free you from managing servers.
00:09:00And it did.
00:09:02But every one of you who shipped a serverless function knows that it came with a catch.
00:09:08Actually, a whole list of them, depending on the vendor.
00:09:13Your function can only run for so long.
00:09:17There were rules on how big your function could be.
00:09:21Network access was often wide open.
00:09:25Good luck getting an FFmpeg to work.
00:09:28Trust me, I've been there.
00:09:30You get one operating system, or in some cases, none.
00:09:34Take it or leave it.
00:09:37There are vendors out there who don't give you real Node.js or Python.
00:09:40There are limitations on how much you can upload to that function.
00:09:45And forget about holding long connections open.
00:09:48No web sockets, no long live streams.
00:09:52Every one of those is a place where the platform just said no.
00:09:58I'm happy to announce that with Fluid Compute 2.0, that list is gone.
00:10:05You can run functions for up to 30 minutes.
00:10:08Functions can be up to 5 gigabytes in size.
00:10:12You get private, secure connections to your own backends.
00:10:17You can install any package.
00:10:19You can pick your operating system.
00:10:22Still, as always, on Vercel, you get real Node and real Python.
00:10:27Large uploads are no problem.
00:10:29And there's full WebSocket support.
00:10:34And, of course, you keep everything you always liked about serverless.
00:10:42Automatic security updates and automated scaling.
00:10:46This is 10 years of innovation, building the compute stack of the future.
00:10:50I'm so proud of the hard work our team has done to give you a world free from trade-offs.
00:10:57So, how do we do it?
00:10:59How do we flip every one of those no's to a yes?
00:11:04Most platforms stitch together a zoo of primitives, bare metal, VMs, containers, WASM, process isolation, V8 isolates, each with their own rules and trade-offs.
00:11:18We made one bet instead.
00:11:20We bet everything on micro VMs and built a platform we call Hive.
00:11:27A micro VM gives you a strong, no-compromised kernel-level isolation between tenants.
00:11:33There are no shortcuts on security.
00:11:36Our bet was that micro VMs could model every kind of compute Vercel runs.
00:11:41Now, that journey started with builds a few years ago, actually.
00:11:45Today, every compute primitive Vercel runs on that one unified stack.
00:11:50That's builds, sandboxes, and functions.
00:11:55Now, a few weeks ago, we brought Docker and OCI images to sandbox.
00:12:03But I just mentioned how functions and sandboxes are now run on one unified stack.
00:12:08So, could I run Docker on Vercel functions?
00:12:12Yes, you can.
00:12:13Today, we're bringing Docker images to the whole platform with Vercel container registry and flexible runtime for fluid compute.
00:12:20Thank you.
00:12:25You heard that right.
00:12:27You can now run Docker images on Vercel and truly ship anything you want.
00:12:32Ruby on Rails, PHP, Java, even .NET.
00:12:36Now, all work natively on Vercel.
00:12:39And like G said earlier, you can run anything on our infrastructure.
00:12:43Frontends, backends, and backend-only services.
00:12:46And now, you can run your own containers.
00:12:49But developers also tell us it's too hard to wire all these pieces together.
00:12:55And so, we fixed it.
00:12:57Today, I'm excited to announce Vercel's services.
00:13:00It's a developer experience you know and love for your full stack applications.
00:13:04You can now develop your frontend and backend together with one command, vcdef.
00:13:11And everything spins up locally.
00:13:13When you push a commit, you get a preview URL for your entire app.
00:13:19Not just the frontend.
00:13:20Even backend-only commits generate a full preview you can test before you ship.
00:13:25And all servers that you deploy can talk to each other privately without ever touching the public internet.
00:13:33You can now run all your microservices on Vercel, and everything just works.
00:13:39Alright.
00:13:41Now, let's talk about the elephant in the room.
00:13:43What might it be?
00:13:45This is all great, but what does it cost?
00:13:51You could throw all of this on a $5 VPS, right?
00:13:55Or as they call it, post-inflation, the $25 VPS.
00:14:00Jokes aside, here's the rebuttal.
00:14:03Fluid compute uses active CPU pricing.
00:14:06You pay the CPU price only when you're actually computing.
00:14:10So, when your agent is sitting there waiting on an LLM, that is zero CPU cost.
00:14:16This is a compute primitive and pricing model that was designed for agents.
00:14:23Now, I know what some of you are thinking, but I can run my own server cheaper.
00:14:27And, to be honest, I was kind of thinking that myself.
00:14:30But can you actually do this?
00:14:33Let's go through this model.
00:14:35Start with us.
00:14:36Fluid is one flat price.
00:14:39It's about 15 cents per active CPU hour.
00:14:43Again, you only pay for the CPU you actually burn.
00:14:47A server you rent is the opposite, right?
00:14:51You pay for the whole machine every hour, whether you use it or not.
00:14:56So, math in our head, renting only wins if you keep that machine very busy.
00:15:03That's just to match our price on raw EC2, you have to run it at 35% utilization every hour, all year.
00:15:12That's your production systems.
00:15:14Possible, for sure.
00:15:16It's your staging systems and the developer sandboxes.
00:15:20Everything together must be above 35% to match fluid.
00:15:23In practice, almost nobody runs that hard.
00:15:29Data for real-world average utilization communities in 2026 is that companies, on average, achieve 8% utilization when considering all those systems.
00:15:40That means you actually use it about one hour in every 12 hours.
00:15:45So, that 15 cents CPU hour in fluid compute, on EC2, it costs you 58 cents.
00:15:52And if you add Fargate or EKS, it's even more.
00:15:56It's the same compute, three or four times the price, and you get the privilege of running it yourself.
00:16:04With active CPU pricing, you never manage any of this.
00:16:08You pay for the CPUs, and that's it.
00:16:13Cool.
00:16:14Let's summarize Flu Compute 2.0.
00:16:17Up to 30 minutes of function duration.
00:16:20Up to 5 gigabytes of code per function.
00:16:23Continued support for VPC peering, VPNs, and static egress IPs.
00:16:29Full Docker image support.
00:16:31Great DX for the cell services.
00:16:33And choose your own OS with continued real support for Node.js, Python, and even .det.
00:16:40Up to 100 megabytes of uploads.
00:16:43Full WebSocket support.
00:16:45And active CPU pricing that beats renting servers in most real-world situations.
00:16:51Vercel is now the most capable compute platform for every type of workload.
00:16:55Front ends, back ends, and agents.
00:16:58Let me bring up Tomo, our chief product officers,
00:17:02to show you what you can build and run on top of it.
00:17:13Hello, everyone.
00:17:14Malta, thank you so much.
00:17:16I'm so happy to be back in New York with all of you.
00:17:19With the latest compute innovations Malta just walked us through,
00:17:23Vercel is now the ideal platform for deploying any type of full-stack application.
00:17:28That includes back ends.
00:17:29It's also the ideal platform for building and deploying agents.
00:17:34Agents need to connect to models, execute complex workflows, and connect to data and apps.
00:17:42And Vercel's agent stack provides all the tools you need to build and ship truly capable agents.
00:17:48Let me walk you through these tools now, starting with how we connect to LLMs using the AI SDK.
00:17:56AI SDK was initially released three years ago, last week, and has been in active development ever since.
00:18:02It's a universal toolkit for building AI applications, frameworks, and agents across any model provider.
00:18:09It's platform-agnostic, and it allows you to generate text, images, audio, and more.
00:18:16With over 16 million NPM downloads every week, it's become the standard TypeScript SDK for accessing any model from any provider.
00:18:24And we also recently released the AI SDK for Python.
00:18:30AI SDK is being used at massive scale by companies all over the world, like Brex.
00:18:36Brex helps businesses with spend management, expenses, corporate cards, and things like that.
00:18:41They build agents that review thousands of transactions in order to audit expenses.
00:18:46Those agents call models, invoke tools, stream results, and parse structured financial data to identify anomalies.
00:18:55And the AI SDK simplifies all of this with a standard abstraction layer.
00:18:59The best part is that you can experiment with different models for different parts of your application,
00:19:05and you can try out new models as soon as they're released, all without needing to make any changes to your product code.
00:19:12Last week, we released our latest major version, AI SDK 7.
00:19:17It adds a ton of new capabilities for working across the entire agent lifecycle.
00:19:22It adds a unified API for reasoning, strongly typed tool and runtime context, support for tool approvals, durability, timeouts, and so much more.
00:19:32But it also enables you to generate videos and build real-time voice apps and agents.
00:19:38Okay, so AI SDK makes building with AI easy, but you still need to manage the connection to the underlying models and providers.
00:19:46That's where our next agent stack primitive comes in, which is the AI Gateway.
00:19:51AI Gateway provides a unified interface for accessing AI models, but it's actually so much more than that.
00:19:57Internally, I've been calling the AI Gateway a token delivery network, and I want to explain what I mean by that.
00:20:03In the early days of the web, there was something called the "Hot Origin Problem."
00:20:08Does anyone remember this?
00:20:10It was actually called the "Hot Origin Problem."
00:20:12Because everyone doesn't agree with me about the naming here, but this is a real thing.
00:20:16Popular sites would suddenly receive spikes in traffic from everywhere, but content lived in only one or a few origin locations.
00:20:24This created overloaded servers, slow downloads, and a pretty unpredictable user experience.
00:20:30The web sort of outgrew the idea that every user should fetch every asset directly from the origin, and so the CDN was born.
00:20:39The CDN became the Internet's performance and reliability fabric, providing intelligent routing, distributed edges, observability, and so, so much more.
00:20:48Similarly, AI use cases have now sort of outgrown the idea that every token should be fetched directly from the model provider.
00:20:57Tokens have actually become a production dependency now.
00:21:01We don't just use them during dev, we use them all the time, and the model labs are becoming the new origins.
00:21:07They're powerful, but they're also expensive, rate-limited, and both geographically and operationally variable.
00:21:15As our friends at the model labs know, this is a really hard problem at scale, and so that's why we built the AI Gateway.
00:21:26It routes around failures, it serves tokens reliably through the same global network that Vercel has run for over a decade.
00:21:34It also simplifies off, centralizes usage and spend tracking with granular observability, and offers zero data retention.
00:21:43But there's another important problem that the AI Gateway solves, which is model choice.
00:21:49Today's agents don't actually use a single model architecture anymore, they use many models from many providers.
00:21:56Our AI Gateway production index recently showed that teams running agents at scale route across 35 different models.
00:22:02So model choice isn't just about experimentation anymore, it's how you actually run AI in production at scale.
00:22:09Surhant is one of the most recognized names in real estate, and they use Vercel's AI Gateway to build an AI app that helps their team with property valuation, marketing, and client communication.
00:22:21They use different models for different tasks based on the need for power or if they're trying to watch cost.
00:22:27Cloud performs complex market analysis, GPT is used to create marketing copy, and Gemini generates images, and all of this happens with a single AI Gateway API key.
00:22:38AI Gateway is already serving over 1 trillion tokens per day across hundreds of models from dozens of providers.
00:22:49It's a huge part of what makes Vercel the open platform for AI.
00:22:55Okay, so Vercel now serves pixels and tokens instantly and reliably, but modern software doesn't really follow a simple request-response model anymore.
00:23:06Agents may need to run for minutes, hours, or even days, and across many, many complex tasks.
00:23:13So that's where the next tool in the agent stack comes in, which is the workflow SDK.
00:23:19Long-running workflows and back-end jobs are nothing new in software, but they're one of the most challenging things to build.
00:23:26Things can go wrong, and they do, and I'm sure many of you have experienced timeouts, dropped state, dropped connections, and more.
00:23:35Without a primitive for durability, you have to kind of hand-stitch together things like retries, state persistence, and this can get really ugly.
00:23:44This is why we built the workflow SDK.
00:23:47We like to say it provides infinite compute durability.
00:23:51It enables you to build long-running apps and agents that automatically suspend, resume, retry, and maintain state with ease.
00:24:00DoorDash uses the workflow SDK to run traditional ETL jobs, keeping data in their app up-to-date.
00:24:06And our customer, Flora, built an entire AI design platform on the workflow SDK to help designers generate visual content at scale.
00:24:15Inside their platform, agents fan out across 50 different image models to generate visual directions from a single creative brief.
00:24:24Workflow SDK checkpoints every step of every agent, and pauses when the job needs human input.
00:24:31And because every failure is automatically retried, the designer never needs to start over.
00:24:36The whole idea behind agents is that they can solve complex tasks across multi-step workflows.
00:24:43And what we've found is that one of the primary ways agents want to accomplish tasks these days is by writing and executing code.
00:24:50That's where our next agent stack primitive comes in, Vercel sandbox.
00:24:55LLMs have gotten really good at producing working code to accomplish a task, but that code is still untrusted, and it can't be run inside the same environment that has access to our production systems.
00:25:07We need a special production-grade environment that's designed for executing untrusted code in a secure and isolated way.
00:25:14Vercel is actually no stranger to this problem of untrusted code execution because of preview deployments and builds.
00:25:22We host over a billion production-grade preview deployments and run over six million builds every single day.
00:25:29Every one of those builds happens inside an isolated micro-VM compute environment.
00:25:34With Vercel sandbox, that same isolated compute environment is now available to you and your agents.
00:25:41It's the safest way to run code you didn't write, and it's perfect for agents.
00:25:46Each sandbox is a fully functional computer with a file system, security boundary, and even full docker support, as Malta just mentioned.
00:25:55And of course, it's all built on top of fluid compute.
00:25:59Vercel sandbox is already being used in production at scale by the best companies in the world, including Notion,
00:26:04Notion.
00:26:05Millions of teams use Notion as their AI workspace, capturing knowledge, answering questions, and moving projects forward.
00:26:12And now, you can extend Notion with custom agents.
00:26:16This enables you to do things like sync CRM data, turn Slack threads into content, and connect to external workflows.
00:26:25But custom agents need code, and that code has to run safely, which is why Notion runs custom agents on Vercel sandbox.
00:26:32Since each agent gets its own general-purpose compute environment, everything outside the sandbox stays protected.
00:26:41Okay, so now we can execute code securely inside long-running workflows.
00:26:45In order to do anything useful, our agents need to access data and tools, and that's where the next layer of the agent stack comes in, starting with Vercel Connect.
00:26:56Vercel Connect is a brand-new building block that allows your agents to securely connect to all the data and systems they need.
00:27:02With Vercel Connect, your agents can access tools like Slack, GitHub, Salesforce, Snowflake, and so many more.
00:27:08And you can even create your own custom connectors with OAuth and API keys.
00:27:13But here's the most important part.
00:27:16Instead of long-lived secrets, every connection uses a short-lived, minimally scoped access token.
00:27:23So, an agent never has standing access to your systems, and it never touches data it shouldn't.
00:27:31This short-lived access token point is really, really important, and minimal, minimal scoped access.
00:27:38Zapier makes it easy for people to build AI automations and workflows.
00:27:41They use Vercel Connect to give their apps private, IP-stable network access to internal AWS resources, like RDS, EKS, and more, via private link and VPC peering.
00:27:53That means their apps can use real, live business data, not just whatever is reachable via a public API.
00:28:01With Vercel Connect, your agents can securely connect to your full range of internal systems.
00:28:05Your CRM, your ERP, your HRIS, data warehouses, and more.
00:28:11But one of the most important connections your agents need is to the tools we use to interact with them.
00:28:16And that's where the next primitive in the agent stack comes in, which is the Chat SDK.
00:28:22Chat SDK provides an elegant abstraction layer, enabling you to interact with agents across dozens of apps.
00:28:28The most powerful agents we've built at Vercel are effectively our coworkers,
00:28:33and we interact with them where work is happening, as it's happening.
00:28:37With just a single line of code, Chat SDK enables your users to interact with agents across tools like Slack, Teams,
00:28:44Google Chat, Discord, GitHub, Linear, Telegram, WhatsApp, and so many more.
00:28:50NanoClaw helps companies run AI agents, and they built their platform on top of the Chat SDK.
00:28:55It's a single code base, but it delivers agents across 15 messaging apps.
00:29:02So that's the Vercel agent stack.
00:29:04It's a singular set of end-to-end capabilities covering everything needed to build and ship agents to production.
00:29:10We've taken everything we've learned building agents over the past few years and turned those
00:29:14learnings into best-in-class primitives that work at Vercel scale.
00:29:19We love this stack, and our customers do too.
00:29:22It's a powerful stack that fills a real gap in the ecosystem.
00:29:26But there's still quite a bit of complexity.
00:29:29Each of these primitives still needs to be wired up into a single cohesive agent.
00:29:34What if they didn't?
00:29:37Vercel, we don't want to just build the world's most powerful agents or enable you to build the
00:29:41world's most powerful agents.
00:29:43We want those agents to have craft.
00:29:45We want building them to be simple.
00:29:47And that means we also have to provide you with the fastest and easiest way to stitch all of those
00:29:52primitives together.
00:29:54To make that possible, we built the newest member of the agent stack family, Eve.
00:30:01Thank you.
00:30:05Eve is our platform-agnostic framework for building complete end-to-end production agents.
00:30:11It's opinionated based on everything we've learned over the past two years, but critically,
00:30:16it's also completely open source and modular.
00:30:20It's built to work seamlessly with our infrastructure with high cohesion between Vercel and Eve,
00:30:26but it's completely customizable so you can make it your own and run it anywhere you'd like.
00:30:31With Eve, every agent is just a directory laid out in a truly intuitive way.
00:30:37We like to call it the Next.js for agents.
00:30:40I have a lot more to share about Eve, but first, I want to welcome Char to the stage
00:30:44to walk us through a demo.
00:30:54Thanks, Tomo.
00:30:55Over the last year, we've taken everything we've learned about building agents and repackaged it
00:31:00into Eve.
00:31:01Let me show you how it works.
00:31:04I'll get started with a single command in my terminal.
00:31:07This command will scaffold an agent directory, install dependencies, and start an interactive chat
00:31:13session with the agent.
00:31:16Next, I'll configure our model provider.
00:31:18I'll select Vercel AI Gateway, select the Vercel team,
00:31:25links to a project in that team, and we're done.
00:31:31Now, let's test the agent.
00:31:33Who are you?
00:31:37Okay, that was fast.
00:31:39I just built a fully functional agent in less than a minute.
00:31:43Now, this agent runs with just two files, agent.ts, which configures the agent's model,
00:31:50and instruction.md, that defines the agent's identity.
00:31:55And that simplicity is what makes it so easy to build an agent with Eve.
00:31:59Now, let's build a real use case, a go-to-market sales agent that processes call transcripts and
00:32:05updates Salesforce in linear.
00:32:07Let's build that agent with Eve.
00:32:09First, I'll give the agent an identity.
00:32:12In the instructions markdown, I'll describe a go-to-market assistant that can create linear
00:32:17issues and update Salesforce opportunities based on call transcripts.
00:32:23And just like that, the agents have an updated charter.
00:32:29Next, the agent needs to create issues, so I'll connect it to linear.
00:32:33And connections are how Eve plugs into external MCP and open API servers, and they go in the
00:32:39Connections folder.
00:32:45I'll define a MCP client connection and use the linear connector in my Vercel team,
00:32:52and I can limit the linear MCP tools the agent can use.
00:32:58I also want the agent to know how to create linear issues, so I'll add a skill.
00:33:02And skills go to the Skills folder.
00:33:08This skill will tell the agent it should look for feature requests in the transcript
00:33:13and separate linear issues by topic.
00:33:17Now, let's test the agent.
00:33:19It should be able to create linear issues.
00:33:26Great.
00:33:27Issue was created, and we should also see it in linear.
00:33:35Next, the agent needs to update Salesforce opportunities.
00:33:38We want to use the Salesforce CLI to make updates and the commands to be executed safely in isolation.
00:33:44I'll configure sandbox.ts.
00:33:48This sets up a private VM for the agent to work in.
00:33:51And the sandbox is only allowed to talk to Salesforce and nothing else.
00:33:57I'll give the agent a Salesforce tool in the tools folder.
00:34:04This will run the Salesforce CLI inside that sandbox.
00:34:08And Salesforce updates need approval from someone on the team, so I'll also give it human in the loop.
00:34:12So, it always requires an approval from a human.
00:34:16And lastly, the tool execution function that will only update Salesforce opportunity.
00:34:23Now, let's test the Salesforce step.
00:34:24I'll ask it to update Salesforce opportunity.
00:34:28That's human in the loop.
00:34:33And it's updated.
00:34:34And if you go to Salesforce refresh, we should see the updated status.
00:34:41Great.
00:34:41Eve also comes with evals support out of the box.
00:34:45So, I can run the agent against real sessions and grade the results.
00:34:50Eve discovers evals under the evals folder.
00:35:00I'll set a config and specify a model to judge the evals.
00:35:05Then create a transcript with two different requests.
00:35:08And check that the agent opened two issues, split them by topic, and didn't touch Salesforce on its own.
00:35:16Now, let's run the eval.
00:35:25Great.
00:35:26It passed.
00:35:26Now, a prompt change can quietly break the agent.
00:35:30We are now ready to make the agent available in Slack for the sales team.
00:35:34So, back to TUI, I'll run slash channels.
00:35:42Add Slack.
00:35:44Yes, I want a Slack bot.
00:35:50Create a Slack connector.
00:35:51This will install the Slack bot in the workspace.
00:35:59And deploy to Vercel.
00:36:04Let's test it end to end.
00:36:06I'll give it an example transcript in Slack.
00:36:10Tag the agent.
00:36:16It'll ask me to approve the change.
00:36:21And that's it.
00:36:22We built a production-ready agent that our team can use.
00:36:25But we're not done yet.
00:36:27This agent is also fully observable.
00:36:29So, let's take a look at that conversation we just had with it.
00:36:32In the Vercel dashboard, click on observability, agent runs, and here's the conversation.
00:36:40I can see the entire conversation history.
00:36:43The token usage, the length of run, the inputs, outputs, and the agent reasoning.
00:36:49I can even drill down into each of the tool calls the agent made.
00:36:55And lastly, EVE is open source.
00:36:57You own your code, you can swap the primitives, and Vercel is the fastest way to deploy it,
00:37:02but it's not the only way.
00:37:03And that's EVE.
00:37:04That's EVE.
00:37:05We just built a production-ready agent in five minutes.
00:37:08Back to you, Tom.
00:37:17We launched EVE two weeks ago at SHIP in London, and it's been incredible to see what everyone has built already.
00:37:25Over 1,500 teams have deployed over 40,000 agents on EVE, and it's got over 250,000 NPM downloads in the last week.
00:37:36But developers aren't just experimenting with EVE, they're running agents in production at scale.
00:37:42Clay built a data analyst agent on EVE, Solomon built an incident response agent on EVE, and
00:37:49Bask Health built an agent that lets non-technical users design and preview product features.
00:37:55Here's a quote from the team at Bask Health that I'm really proud of.
00:37:58We've gone through a few different iterations over the last six months, but EVE has been the most
00:38:02batteries-included experience.
00:38:04I can't believe how quickly we got up and running.
00:38:07This was exactly our goal.
00:38:11Now, we've built a ton of agents ourselves, as we've mentioned a few times, but one that I'm
00:38:17particularly proud of, we gave the Vercel name.
00:38:20We just call it Vercel Agent.
00:38:23I like to describe it as an intelligence layer for Vercel, because it can help you do almost
00:38:28anything on the platform.
00:38:30We built a beautiful web UI so you can chat with Vercel Agent directly from the dashboard and ask it
00:38:35almost anything.
00:38:37For example, you can ask it to fix the 500s that are showing up in your logs, and it'll perform an
00:38:43investigation.
00:38:44You can ask it to find accessibility issues in your project, and it'll run a review and open a pull
00:38:49request.
00:38:51Or you can simply ask it to fix your build, and it'll read the deployment logs, find the failing config,
00:38:57validate the fix in a sandbox, and redeploy with your approval.
00:39:02But what I love most is that the Vercel Agent is not just another chatbot.
00:39:06It sees your app running in production and proactively monitors your infrastructure on your behalf.
00:39:11Let me give you an example.
00:39:13Let's say you get an alert about a partial outage.
00:39:16When you log into the dashboard and click on the alert, Vercel Agent will have already run an
00:39:20investigation on the anomaly.
00:39:23Here we can see it found that the API key object is undefined at runtime, and errors appeared after the
00:39:29last deployment a few minutes ago.
00:39:32It recommended an instant rollback to the previous release, which, when in doubt, is always the right
00:39:36action to take.
00:39:38So this one's an easy one to approve.
00:39:41With permission, Vercel Agent rolled back to the last production deployment before the 500s showed up.
00:39:47And after the rollback was successful, our app is back online and Vercel Agent can kick off work on a proper
00:39:53fix for the error.
00:39:55Now, the first question every CTO asks is, is it safe to let the agent do that?
00:40:02And this is the right question, because most agents inherit their users' permissions.
00:40:07They run as you do and can do everything you can do.
00:40:11So a single bad prompt can have extremely wide blast radius.
00:40:16But Vercel Agent has a first-of-its-kind permission model.
00:40:19It has its own identity, and it's read-only by default.
00:40:23First, Vercel Agent presents you with a plan, not an unknown number of approvals.
00:40:29And instead of asking, can I do this, can I do this, you approve the entire plan up front.
00:40:34So you know what the outcome is going to be before Vercel Agent does any work.
00:40:39Then it's granted permission scoped exactly to that plan and nothing more.
00:40:44Every agent runs in a sandbox before it touches production.
00:40:47And anything that changes your production app waits for your approval.
00:40:51So it never has more access than the task needs, and it never stops to ask you again.
00:40:57Vercel Agent is just one of the many agents that we've built.
00:41:01And I'm excited to welcome Gene, our COO, to the stage to tell you about all of the agents
00:41:05we've built to transform our go-to-market motion.
00:41:08Thank you.
00:41:18Thanks, Tomo.
00:41:20Vercel Agent shows you what's possible when you build agents the right way.
00:41:25And every company in this room is going to build an agent just like it.
00:41:30There are two types of people in this room hearing that. Some of you are sitting there thinking,
00:41:34"Let's go. I am shipping tonight." And then there are the CIOs and the CTOs in the room who are
00:41:41thinking, "Oh, no!" Because you can already feel what's coming. Shadow agents writing to systems with no audit trail.
00:41:49AI bot user closing your tickets. Spend you can't explain.
00:41:54Both of you are right. Building agents is easier than you think and way harder than you think.
00:42:03I'm going to tell you about that tension and what we've learned at Vercel from living it.
00:42:08Drew Bredvick, who works for me, is head of go-to-market engineering.
00:42:12In June of 2025, he had the sexiest job in the world.
00:42:17His mandate? Build the agents that would transform how Vercel goes to market.
00:42:23It worked. A year later, agents are part of our daily workflow, running at scale across our entire
00:42:29go-to-market organization. You did a great job, Drew. So today, I'm handing you a pager.
00:42:37Let me tell you why. What we learned was that agents are free. As in, free puppies.
00:42:44Everyone loves puppies. But they pee on your floor. They eat your furniture. And you can't go on vacation.
00:42:52Agents are free because anyone can prompt COD. But agents are also software. And we all know that
00:42:58software is never done. Someone has to maintain them, update models, and build new features.
00:43:05Building hundreds of agents taught us hard lessons. First, we saw the same problem solved over and over.
00:43:13Multiple agents needed to connect to the same internal systems. Each team built their own integrations
00:43:19from Slack. Second, each agent was reading from different knowledge bases. Our team would ask the
00:43:26same question and get different answers. Third, we had no visibility. No one knew how many agents
00:43:33existed, who built them, or what data they touched. Fourth, that lack of visibility also meant
00:43:41adoption chaos. I'm on Slack. I type at. 500 agents pop up. I don't know what any of them do.
00:43:49And last, we also learned that chat isn't all that you need. We started with the idea that Slack was the
00:43:55universal interface. And that was wrong. The agents that actually got used also had front ends for
00:44:03permissions, for visualizing data, for workflows, and for keeping humans in the loop. We learned that our agents
00:44:10had to work on day one and day 100. And I'm happy to tell you that they do. We run over 100 agents in
00:44:20production at Vercel. And they're part of how we operate every single day. I want to tell you about 10 of the most
00:44:28important ones. And the order matters. We started with the most obvious use cases and worked our way towards
00:44:35the agents that changed our internal processes and transformed how we operate the go-to-market team.
00:44:42First, Vertex is our customer support agent. It resolves over 91% of support tickets across the help center,
00:44:52Slack, and DocsChat. Second, Deal 1 is our Deal Intelligence agent. It listens to every sales call,
00:45:01coaches our sales reps in Slack, and runs a postmortem on every lost deal. The Deal 1 MCP has been called
00:45:0817,000 times in the last month. Next, Draft 0 is our content agent. It writes the first draft of every
00:45:18blog post, change log, and customer story we publish. A0 is our AEO agent. It tracks how Vercel's brand and
00:45:27content shows up across AI search. Every day, it runs hundreds of prompts across dozens of coding agents.
00:45:36Revoa is our Salesforce update agent. It publishes critical record changes to Salesforce with a human in
00:45:43the loop. It saves nine hours of time for our RevOps team every single day. And Penny is our finance and
00:45:52Ops agent. It has access to our billing platform, payment provider, and monitoring systems. It saves our
00:45:59finance and on-call engineering teams hours of answering billing questions. And then there's V. V is our
00:46:07routing agent. It routes requests to all of our other internal agents. Remember that adoption problem?
00:46:14We still have 100 agents. But V is the front door to all of them. Ask V a question and it picks the
00:46:20right agent for the job. The last three I'm going to show you are the ones I want to spend real time on
00:46:27the job. Because every one of you is going to want to build them. First, D0, our data agent. D0 gives the
00:46:35entire company 24/7 on-demand analytics and data science. Anyone at Vercel, engineers, AEs, finance,
00:46:43support, can run analytics on data from our warehouse without filing a ticket or waiting on the data team.
00:46:50Users can ask simple questions like how many leads we got from this campaign. And D0 writes and runs SQL.
00:46:57But D0 is also a data scientist. If users need statistical analysis done, it spins up a sandbox
00:47:06and runs Python to generate reports. D0 is the most used internal tool at Vercel. It handles 30,000
00:47:14questions per month. And it's safe at scale. V0 doesn't run in God mode. Every query is scoped to the
00:47:22user's permissions. If you can't see a table in Snowflake, D0 can't show it to you either. Under the
00:47:28hood, D0 needed a semantic layer. And even though you can ask questions in Slack, it needed a UI so
00:47:35people could explore charts and drill into data. Next, Athena, our sales cockpit. Salesforce announced
00:47:44headless. We've been running on it for months. Athena picks accounts, plans outreach, tracks signals, and
00:47:51the weekly motion for every AE at Vercel. Shortly after we went live, pipeline coverage nearly doubled.
00:47:59Every AE uses it every single day. Under the hood, Athena needed the same semantic layer D0 needed,
00:48:07plus durable workflows and secure connections, plus a UI, because GTM agents are more effective with pixels
00:48:14and buttons than just a Slack channel. Third, our lead agent, an autonomous SDR. We trained lead agent on
00:48:23the playbook of our best SDR, and now it runs that playbook 24/7. You may have seen the headline last
00:48:30year. We took 10 SDRs down to one. That was lead agent. We redeployed those nine reps into bigger
00:48:38roles and raised our quota. We've seen a 32X ROI, and it costs $5,000 a year to run. It performs at the
00:48:4890th percentile of our reps, and one engineer maintains it part-time. Under the hood, lead agent runs on the
00:48:56stack Tomo just showed you. AI SDK, workflows SDK, chat SDK. And it's open source. So, you can go build your own today.
00:49:07Building these agents was easy. Because we used Next.js and our agent framework Eve. And running them was
00:49:14never a problem because they run on Vercel. Remember when I said that building agents is way harder than you
00:49:20think? The hard part is everything around them. Who can access them, how they authenticate, what data they
00:49:28can touch, and providing all of this to your security team. So, we spent the last year building a platform
00:49:35that makes that easy. And today, we're making it available to you. I'm excited to announce Vercel for
00:49:43enterprise apps and agents. It's the Vercel developer experience you love for everyone at your company.
00:49:51With identity and access built in. An option to run it in your own AWS tenant. Like all of our products,
00:50:00it's portable because it's framework and model agnostic. We built it for ourselves first, and now it's the
00:50:06platform you can build on. Let me show you the most important pieces. I said earlier that anyone can
00:50:15prompt Claude. The reality is that your employees are already doing this, whether you know it or not.
00:50:21That's called shadow IT. And AI has already caused major data breaches in the enterprise. So, even if you
00:50:28control who builds AI, you still have to limit access to the apps and agents they build.
00:50:35That's where Vercel Passport comes in. Whether you're using V0, Claude code, codex, everything your team
00:50:43builds stays internal. Governed by the identity provider you already run. Like Okta or Microsoft
00:50:50Entra. And it goes deeper than the front door. Passport issues a token tied to each user's identity
00:50:58and carries it through to the backends and systems your apps and agents connect to. At Vercel,
00:51:05we use V0 to build data apps and reports powered by our native Snowflake integration. This allows us to
00:51:11build apps that have direct access to our Snowflake data. But Passport doesn't just protect access to
00:51:17the application. It also limits the data users can see in the application based on permissions provided by
00:51:25our IDP. I'll give you another example from our customer Mercor. Their security team built an internal agent
00:51:32that understands their entire code base. You can ask it questions like, which of our public routes
00:51:38connect to a database? And it will pull answers across every repository. It will even kick off fixes.
00:51:44But what makes it interesting is not just people asking the agent questions. Other agents query it, too.
00:51:52So, they had a hard problem. They wanted employees and agents to be able to use it. And they wanted to
00:51:57control access. But they didn't want to hand build and manage authentication for it. They implemented
00:52:04Passport and Connect in under a week. And now the agent runs on Vercel behind Okta, the IDP they already use.
00:52:12The right people and agents get in. Scoped automatically with no homegrown auth. Passport
00:52:20controls who gets into your apps. But when everyone is shipping, you also need the ability to look
00:52:25across deployments and see if there are security risks. That's why we built the security dashboard.
00:52:32It's one place to see the security posture of every account and project on Vercel. It flags
00:52:39misconfigurations that can lead to breaches, shows you members without MFA, alerts you about secrets shared
00:52:46across projects, and identified long-lived credentials when short ones will do. The best part is that it
00:52:53actually shows you how to fix each problem. I'm excited to tell you that today, all of those deployments
00:53:01can run in your own AWS tenant. You heard that right. You can run Vercel functions in your own AWS account.
00:53:10Whatever you build with agents stays inside your cloud. Everything I've shown you today runs on
00:53:17Vercel's core platform. It's the same platform the world's best teams have trusted for over a decade
00:53:23for our developer experience, previews, firewall, end-to-end observability, and features like
00:53:29instant rollback. These new capabilities aren't bolted on. They're a part of the Vercel you already
00:53:35trust. The teams who ship fast, securely, at scale are the teams who will win. We built enterprise apps and
00:53:44agents, so that can be you. Gee, back to you to wrap up. Thanks, Gene. Today, we showed you that Vercel
00:54:00is a platform where you can build and run everything. Any kind of website, app, or agent. Your backend
00:54:10frameworks can run at scale. With Vercel services, you can develop, preview, and ship backends and
00:54:18frontends together. Fluid compute, as Malta showed us, lets you run any type of workload, from instant
00:54:28serverless functions to backend agent jobs that take hours, days, months to complete, and you can even bring
00:54:36your own container. With Eve, your team can build a production agent in minutes. Vercel Passport makes
00:54:47sure your internal agents and apps stay internal, behind your IDP. And you can run all of this
00:54:57in the security beauty of your AWS account. It's beautiful. And you can start today. Ask your coding
00:55:06agent to install the Vercel plugin. Then, you can build anything. A website, an app, an agent, and ship it
00:55:16at global scale. And Vercel agent, of course, will keep an eye on production.
00:55:23This is agentic infrastructure. We can't wait to see what you all ship next. Thank you very much.