Ship 26 London - Keynote

VVercel
Computing/SoftwareSmall Business/StartupsManagementInternet Technology

Transcript

00:00:00- Hi, everyone.
00:00:11Thank you, thank you.
00:00:12Hi, everybody, it's so great to be here.
00:00:15Since SHIP last year, the world has changed.
00:00:20We're writing less code by hand,
00:00:22but more ideas are coming to life than ever before.
00:00:27Six months ago, less than 3% of Vercel's deployments
00:00:31were triggered by coding agents.
00:00:33Since then, that number has grown 17 times,
00:00:38and over half of deployments on Vercel now come from agents.
00:00:43But the bigger shift is what those agents are deploying.
00:00:48Since the beginning of the year,
00:00:50agentic workloads on Vercel have doubled.
00:00:53Volume on AI Gateway grew from 2 trillion
00:00:57to 20 trillion tokens per month.
00:01:01And the shift we're seeing is that we're using agents
00:01:04to deploy software that can think.
00:01:07We're shipping agentic sites and apps.
00:01:11We're using agents to ship agents.
00:01:14This is a profound change, so I want to go back in time
00:01:17and show you how we got here.
00:01:20Because our new world is agentic.
00:01:23But it all started with websites.
00:01:27I remember building simple sites with HTML, CSS, and JavaScript
00:01:31as a kid, as one does.
00:01:34I was so excited to put my creations online.
00:01:37And that's what the web is all about.
00:01:39Sharing our ideas with the world.
00:01:44We built the first version of Vercel to make that easy for anyone.
00:01:48It was infrastructure for pages, caching, and global content delivery.
00:01:54Today, we serve the fastest, most beautiful websites in the world
00:01:59for companies like Co-op, Paul Smith, and Heli Hansen.
00:02:04But the world is dynamic.
00:02:06And as the web grew, sites turned into data-powered applications.
00:02:11So Vercel built infrastructure for servers, APIs, and databases.
00:02:17Now, sum up TripAdvisor and your very own London Northeastern Railway
00:02:22run apps at scale on Vercel.
00:02:26For over a decade, we've been building Vercel into a cloud
00:02:30where you can run everything.
00:02:32Our vision is to build a true, full-stack platform.
00:02:37And in the last year, we've shipped some huge steps towards that vision.
00:02:42We brought on some of the best, finest Python developers in the world
00:02:47to extend our core infrastructure for backend frameworks.
00:02:51You can now run backends like FastAPI, Flask, Express,
00:02:56and HONO at scale, right on Vercel.
00:03:00We built out a complete agentic layer.
00:03:04You can now run longer-duration functions on Fluid Compute,
00:03:07make workflows durable,
00:03:09and spin up secure sandboxes.
00:03:13You can host your MCP servers
00:03:14and make your app available to any agent.
00:03:18And we added the best databases in the world to our platform.
00:03:23You can manage Amazon Aurora, Aurora D-SQL, and DynamoDB,
00:03:28along with hundreds of high-quality integrations,
00:03:31like OpenSearch, all through Vercel.
00:03:34This means you can run any architecture you want on our infrastructure.
00:03:40Right here in London, Octopus Energy runs a Next.js frontend
00:03:44with a Python backend,
00:03:46helping millions in the UK monitor their power usage.
00:03:50So, in addition to frontends,
00:03:53you can host backend-only services on Vercel,
00:03:57like REST APIs written in TypeScript and Python.
00:04:01You can deploy workflows that handle long-running asynchronous work,
00:04:05even with a human in the loop.
00:04:07And you can host queues that process billions of messages per month.
00:04:13But developers tell us it's still too hard
00:04:16to wire all of those pieces together.
00:04:20So, we fixed it.
00:04:23Today, I'm excited to announce Vercel Services.
00:04:28It's a developer experience you know and love
00:04:31for your full-stack application.
00:04:34You can now develop frontend and backend together
00:04:38with one command, vcdev.
00:04:41Everything spins up locally.
00:04:43And when you push a commit,
00:04:46you get a preview URL for your entire application,
00:04:49not just the frontend.
00:04:51Even backend-only commits
00:04:53generate a full preview you can test before you ship.
00:04:58And all of the services you deploy can talk to each other privately
00:05:02without ever touching the yucky public internet.
00:05:07You can now run all of your microservices,
00:05:09I said microservices,
00:05:11microservices on Vercel,
00:05:13and everything just works.
00:05:16But in the agent era,
00:05:18these services are becoming autonomous.
00:05:21Websites and apps used to respond to user input with logic.
00:05:27But now,
00:05:29they have agents inside
00:05:30that can understand intent
00:05:31and take autonomous action.
00:05:35Every new generation of software
00:05:37demands a new generation of infrastructure.
00:05:41And the agent era is no different.
00:05:44Agentic software needs agentic infrastructure.
00:05:48And that's exactly what we're building at Vercel.
00:05:51Agentic infrastructure has three parts.
00:05:55One,
00:05:56Vercel is where coding agents deploy software.
00:06:01When you ask Cloud Code or Codex where to deploy,
00:06:04you get Vercel.
00:06:05Because Vercel is built for the way agents work.
00:06:10Second,
00:06:11Vercel is where you build and deploy your own agents.
00:06:15We give you every tool you need
00:06:17to build and run apps and agents
00:06:20in production securely at scale
00:06:22on one platform.
00:06:25And third,
00:06:26Vercel itself is automated by agents.
00:06:31Vercel runs your apps in production, of course,
00:06:34handling traffic, traces, observability, and anomalies.
00:06:39And so that data gives our agents
00:06:41the context they need to investigate autonomously
00:06:44and then service port requests,
00:06:47not just alerts.
00:06:49For context,
00:06:51teams using Cloud Code
00:06:52deploy Vercel five times more frequently
00:06:55than teams who don't.
00:06:57That's because we set the standard
00:06:59for developer experience,
00:07:01and now we're doing the same
00:07:02for the agents that those developers use.
00:07:06Coding agents love Vercel
00:07:07because we give them everything they need.
00:07:11In fact,
00:07:11when you ask your agent
00:07:13to verify its work,
00:07:14it needs a live URL it can test, right?
00:07:18So Vercel gives every deployment
00:07:20a dedicated, secure, review environment.
00:07:24When you ask your agents
00:07:26to ship experiments,
00:07:27it needs to roll them out safely.
00:07:31So Vercel gives every experiment
00:07:33a feature flag
00:07:34and gives you control
00:07:36and confidence
00:07:37with instant rollbacks.
00:07:40And you don't want your agent
00:07:42wasting time
00:07:43clicking around a dashboard.
00:07:46An agent is most efficient
00:07:48when every part of the platform
00:07:50is available
00:07:51in its own language.
00:07:54And that's what Vercel gives it
00:07:55through our CLI,
00:07:57API,
00:07:58and MCP.
00:08:01This year,
00:08:02an engineer at Meta
00:08:03opened Claude code,
00:08:05as one does,
00:08:06and asked you to build
00:08:07that internal tool.
00:08:08And when it was sent to test it,
00:08:10he asked Claude
00:08:11where to deploy.
00:08:13Lo and behold,
00:08:14pushed you to Vercel.
00:08:16A week later,
00:08:17everyone on his team
00:08:18was deployed to Vercel.
00:08:20And within a month,
00:08:21Vercel was the go-to platform
00:08:22for Meta Superintelligence Labs.
00:08:26Meta.ai,
00:08:27Meta's Frontier AI product,
00:08:29was born on Vercel.
00:08:32And that happened
00:08:33even though Meta
00:08:34had already built
00:08:35their own deployment platform
00:08:36for decades.
00:08:38But even the most powerful
00:08:39infrastructure
00:08:40doesn't natively speak
00:08:42the language of agents.
00:08:44Vercel does.
00:08:46Agentic infrastructure
00:08:47is the platform we built,
00:08:49but what matters the most
00:08:50are the products
00:08:51you build
00:08:52on top of it.
00:08:53So I want to invite
00:08:54Tom O'Kino
00:08:55to the stage
00:08:56to show you how.
00:08:57Thank you.
00:09:06Thank you, G.
00:09:07Thank you so much.
00:09:08Everyone,
00:09:09it's so great to see you.
00:09:10It is so great
00:09:11to be back in London.
00:09:13G just told you
00:09:14why coding agents
00:09:15love Vercel's infrastructure,
00:09:17why they choose
00:09:18to deploy
00:09:18to Vercel's
00:09:19agentic infrastructure.
00:09:20But what about
00:09:21when you want
00:09:21to build your own agents?
00:09:25Vercel's agent stack
00:09:26gives you all the tools
00:09:27you need
00:09:28to create and ship
00:09:29your own agents.
00:09:31Agents need context,
00:09:33agents need to connect
00:09:34to models,
00:09:35execute complex workflows,
00:09:37and connect to data
00:09:38and apps.
00:09:40Let me walk you
00:09:40through these tools now,
00:09:41starting with how
00:09:42to connect to models.
00:09:44First up,
00:09:44we have the AI SDK.
00:09:46Have any of you
00:09:46used the AI SDK?
00:09:48Show of hands.
00:09:49Lovely.
00:09:50AI SDK was initially
00:09:51released three years
00:09:52ago this week,
00:09:53and it's been
00:09:53in active development
00:09:54ever since.
00:09:56It's the universal toolkit
00:09:57for building AI frameworks,
00:09:59apps,
00:09:59and agents.
00:10:01It's platform,
00:10:02framework,
00:10:03and model agnostic,
00:10:04and it allows you
00:10:05to generate text,
00:10:06images, speech,
00:10:07video,
00:10:08and more.
00:10:09It also makes it
00:10:10trivial to add things
00:10:11like streaming
00:10:11and fallbacks
00:10:12to your projects.
00:10:14The AI SDK
00:10:15has now downloaded
00:10:16over 15 million times
00:10:18per week
00:10:19and has become
00:10:20the standard way
00:10:20to access any model
00:10:22from any provider
00:10:24in both TypeScript
00:10:25and Python.
00:10:26Shout out to the AI SDK
00:10:27for Python.
00:10:29It's being used
00:10:29at massive scale
00:10:30by companies
00:10:31all over the world,
00:10:32like Brex.
00:10:34Brex helps businesses
00:10:35like Vercel
00:10:36with corporate cards
00:10:36and spend management.
00:10:38They run expense audit
00:10:40agents that review
00:10:40thousands of transactions
00:10:42at a time.
00:10:43Those agents
00:10:44call models,
00:10:45invoke tools
00:10:46mid-generation,
00:10:48stream results
00:10:48as they come back,
00:10:50and parse structured
00:10:51financial data
00:10:52to surface anomalies.
00:10:54The AI SDK
00:10:55simplifies all of this
00:10:57and provides
00:10:57a standard abstraction layer.
00:11:00One of the things
00:11:01we love
00:11:01is that you can experiment
00:11:02with different models
00:11:03for different parts
00:11:04of your application
00:11:04and try out new models
00:11:06as they're released,
00:11:08all without needing
00:11:09to make any changes
00:11:09to your product code.
00:11:11you get to focus
00:11:12on building your features
00:11:13and the AI SDK
00:11:14handles the rest.
00:11:17Okay, so,
00:11:17AI SDK makes integration
00:11:19really easy,
00:11:20but you still have
00:11:20to actually manage
00:11:21the connection
00:11:22to the underlying providers
00:11:23and to the models
00:11:24that you want to use.
00:11:26And that's where
00:11:26the next tool
00:11:27in our agent stack
00:11:28comes in,
00:11:30the AI gateway.
00:11:32AI gateway provides
00:11:33a unified interface
00:11:34for accessing AI models,
00:11:35but it's actually
00:11:36so much more than that.
00:11:38Internally,
00:11:38we've been calling
00:11:39the AI gateway
00:11:39a token delivery network,
00:11:42and I want to explain
00:11:42what I mean by that.
00:11:44In the early days
00:11:45of the web,
00:11:46we had something called
00:11:47the hot origin problem.
00:11:50Popular sites
00:11:50would suddenly receive
00:11:51spikes in traffic
00:11:52from everywhere,
00:11:54but content lived
00:11:55in only one
00:11:56or a few origin locations.
00:11:58This created
00:11:59overloaded servers,
00:12:00slow downloads,
00:12:01and an unpredictable
00:12:02user experience.
00:12:04The web sort of
00:12:04outgrew the idea
00:12:05that every user
00:12:06should fetch assets
00:12:07directly from the origin.
00:12:10And so,
00:12:10the CDN was born.
00:12:12The CDN became
00:12:13the internet's performance
00:12:14and reliability fabric.
00:12:17Distributed edges,
00:12:18intelligent routing,
00:12:20failover,
00:12:21observability,
00:12:22centralized policies,
00:12:24and more.
00:12:26AI use cases
00:12:26have now outgrown
00:12:27the idea
00:12:28that every token
00:12:29should be fetched
00:12:29directly from
00:12:30the model provider.
00:12:32Tokens have become
00:12:33a production dependency,
00:12:35and model labs
00:12:35are the new origins.
00:12:37They're powerful,
00:12:38for sure,
00:12:39but they're also expensive,
00:12:40rate-limited,
00:12:41and geographically
00:12:42and operationally variable.
00:12:44And as our friends
00:12:45at the model labs know,
00:12:46this is a really hard
00:12:47problem at scale.
00:12:49So, that's why
00:12:50we built the AI Gateway.
00:12:52It serves tokens
00:12:53through the same global network
00:12:54that Vercel has run
00:12:55for over a decade.
00:12:56It routes around failures,
00:12:58simplifies off,
00:13:00and centralizes usage
00:13:01with things like spend tracking
00:13:02and granular observability.
00:13:05And, of course,
00:13:06it offers zero data retention,
00:13:08which is particularly important
00:13:09to our enterprise customers.
00:13:11AI Gateway lets teams
00:13:12operate AI traffic
00:13:13as infrastructure
00:13:14instead of bespoke
00:13:16provider integrations.
00:13:19Now, serving tokens reliably
00:13:20is one of the important problems
00:13:22that the Gateway solves,
00:13:23but the other is about
00:13:24model choice.
00:13:26Today's agents
00:13:26don't actually use
00:13:27a single-model architecture.
00:13:29They use many models
00:13:30from many providers.
00:13:33In fact,
00:13:33our AI Gateway production index
00:13:35showed that teams
00:13:35running agents at scale
00:13:37route across 35 different models.
00:13:39That means routing
00:13:40is no longer a feature.
00:13:42It's actually how you run
00:13:43AI in production.
00:13:45One of our customers,
00:13:46AKQA,
00:13:47built a chat GPT app
00:13:49for Starbucks
00:13:49using Next.js
00:13:51and the AI Gateway.
00:13:53It's an interactive bot
00:13:54that helps people discover
00:13:55new drinks they might like.
00:13:57They use different models
00:13:58for different tasks
00:13:59to balance power and cost.
00:14:02Gemini Flash
00:14:03handles retrieval
00:14:04and summarization
00:14:05when someone asks a question.
00:14:07Then recommendations
00:14:08are handled by GPT-5.
00:14:10And even more models
00:14:11are used to generate recipes,
00:14:13all with a single
00:14:14AI Gateway API key.
00:14:17AI Gateway is serving
00:14:18over 1 trillion tokens per day.
00:14:22It gives developers
00:14:23simplified access
00:14:24to hundreds of models
00:14:25from dozens of providers
00:14:26and provides powerful observability.
00:14:29It's actually a huge part
00:14:30of what makes Vercel
00:14:31the open platform for AI.
00:14:35Okay, so Vercel serves pixels
00:14:37and tokens instantly
00:14:38and reliably.
00:14:39But modern software
00:14:40doesn't follow
00:14:41a simple request-response
00:14:42model anymore.
00:14:44Agents may need
00:14:45to run for hours,
00:14:47even days,
00:14:48and across many complex tasks.
00:14:51That's where the next tool
00:14:52in the agent stack
00:14:53comes in,
00:14:54the Workflow SDK.
00:14:57Long-running workflows
00:14:58and backend jobs
00:14:58are nothing new in software,
00:15:00but they're one of the most
00:15:01challenging
00:15:01and annoying things to build.
00:15:03Many things can go wrong
00:15:05and failures do happen.
00:15:07Timeouts,
00:15:08dropped connections,
00:15:09lost state.
00:15:11Without a primitive
00:15:12for durability,
00:15:13you're hand-stitching things
00:15:14like retries
00:15:14and state persistence.
00:15:16It can get really ugly.
00:15:18That's why we built
00:15:19the Workflow SDK.
00:15:20We like to say
00:15:21it provides infinite
00:15:22compute durability.
00:15:23It enables you
00:15:24to build long-running apps
00:15:25and agents
00:15:26that automatically suspend,
00:15:27resume, retry,
00:15:29and maintain state
00:15:30with ease.
00:15:32DoorDash uses
00:15:32the Workflow SDK
00:15:33to run traditional ETL jobs,
00:15:35making sure that all the data
00:15:37in their app
00:15:37is always up to date.
00:15:39And our customer, Flora,
00:15:40sorry for the spoiler
00:15:41a minute ago,
00:15:42built an entire AI design
00:15:44platform on top
00:15:45of the Workflow SDK.
00:15:47Their platform helps designers
00:15:48generate visual content
00:15:49at scale.
00:15:51Inside it,
00:15:52agents fan out
00:15:53across 50 different image models
00:15:54to generate visual directions
00:15:56from a single creative brief.
00:15:59Workflow SDK
00:15:59checkpoints every step
00:16:01of every agent
00:16:02and pauses
00:16:03when the job needs human input.
00:16:05And because every failure
00:16:06is automatically retried,
00:16:08the designer never needs
00:16:09to start over.
00:16:11Now, the whole idea
00:16:12behind agents like this
00:16:13is that they can solve
00:16:14complex tasks
00:16:15across multi-step workflows.
00:16:18And one of the primary ways
00:16:19that agents
00:16:19try to accomplish
00:16:20most tasks
00:16:21and what they love,
00:16:23the way that they love
00:16:24to accomplish most tasks
00:16:25is by writing
00:16:26and executing code.
00:16:27And that's where
00:16:28our next agent stack
00:16:29primitive comes in,
00:16:31Vercel Sandbox.
00:16:33While LLMs have gotten
00:16:34really good
00:16:34at producing working code,
00:16:36that code is still
00:16:37untrusted.
00:16:38And we can't run
00:16:40untrusted code
00:16:41in the same environment
00:16:42that has access
00:16:43to our production systems.
00:16:45We need a special
00:16:45production-grade environment
00:16:47that's designed
00:16:48for executing code
00:16:49in a secure
00:16:50and isolated way.
00:16:53Vercel is actually
00:16:53no stranger
00:16:54to this problem
00:16:54of untrusted code execution
00:16:56because of preview deployments
00:16:58and builds.
00:16:59We host over
00:17:00one billion production-grade
00:17:02preview deployments,
00:17:03and every single day
00:17:04we run over
00:17:05six million builds.
00:17:08Every one of those builds
00:17:09happens inside
00:17:10an isolated
00:17:10micro-VM compute environment.
00:17:13With Vercel Sandbox,
00:17:15the same isolated
00:17:16compute primitive
00:17:17is now available
00:17:18to you and your agents.
00:17:21Vercel Sandbox
00:17:22is built on top
00:17:22of fluid compute.
00:17:24It supports dynamic,
00:17:26real-time workloads
00:17:27for agents,
00:17:28code generation,
00:17:29and developer experimentation.
00:17:31Each sandbox
00:17:32is a fully functional
00:17:33computer with a file system,
00:17:35security boundary,
00:17:36and even full
00:17:37Docker support.
00:17:38When you spin up
00:17:39a sandbox,
00:17:39you can install packages,
00:17:41run containers,
00:17:41and even configure
00:17:42Redis or Postgres
00:17:43as test dependencies.
00:17:45It's already being used
00:17:46in production at scale
00:17:47by the best companies
00:17:48in the world.
00:17:49One company in particular
00:17:51that we use every day
00:17:51at Vercel
00:17:52is Notion.
00:17:54Millions of teams
00:17:55use Notion
00:17:56as their AI workspace,
00:17:57capturing knowledge,
00:17:58answering questions,
00:17:59and moving projects forward.
00:18:01And developers
00:18:01can extend Notion agents
00:18:03with custom code,
00:18:05syncing CRM data,
00:18:06turning Slack threads
00:18:07into content,
00:18:08or connecting
00:18:09to the workflows
00:18:09their teams already rely on.
00:18:11But that code
00:18:12has to run safely.
00:18:14And that's why Notion
00:18:15custom agents
00:18:16run on Vercel Sandboxes.
00:18:19Since each agent
00:18:19gets its own
00:18:20general-purpose
00:18:21compute environment,
00:18:22developers can build
00:18:23almost anything
00:18:24they can imagine,
00:18:25and everything
00:18:25outside the sandbox
00:18:27stays protected.
00:18:29So now our agents
00:18:30can execute code securely
00:18:31inside long-running
00:18:33workflows.
00:18:34But in order
00:18:34to do anything useful,
00:18:35our agents need
00:18:36access to data
00:18:37and tools,
00:18:38the apps that we all
00:18:39use every day.
00:18:40And that's where
00:18:41the next layer
00:18:42of the agent stack
00:18:43comes in,
00:18:44Vercel Connect.
00:18:46Vercel Connect
00:18:47is a brand-new
00:18:47building block
00:18:48that allows your agent
00:18:49to securely connect
00:18:50to all the data
00:18:51and systems they need,
00:18:52including the apps
00:18:53and tools they use
00:18:54to communicate
00:18:55with users
00:18:55via secure,
00:18:57short-lived,
00:18:58and minimally scoped
00:18:59access tokens.
00:19:01I'm excited to invite up
00:19:02Hedi,
00:19:03product lead for Connect,
00:19:04to walk us through
00:19:05a demo of Vercel Connect
00:19:06now.
00:19:08Hedi?
00:19:14Thank you, Tomo.
00:19:16As Tomo mentioned,
00:19:18agents only truly
00:19:20become useful
00:19:21when they have access
00:19:22to the systems
00:19:23and tools
00:19:25that your business uses.
00:19:27Let's say we wanted
00:19:28to build an agent
00:19:29that automatically
00:19:30processes call transcript
00:19:31for our sales team.
00:19:33We might want an agent
00:19:35to also update statuses
00:19:37in Salesforce
00:19:37to reflect what was discussed
00:19:39in the sales call.
00:19:42It could also extract
00:19:43feature requests
00:19:44from the call
00:19:45and then create issues
00:19:47in systems like Linear.
00:19:49And of course,
00:19:50our sales team
00:19:51wants to interact
00:19:52with their agents
00:19:52from Slack.
00:19:54This sounds like
00:19:55a basic agent
00:19:56for processing text,
00:19:58but it's actually
00:19:59a complex workflow
00:20:00with real security
00:20:01implication.
00:20:03When you build
00:20:04custom connections
00:20:05for your agents
00:20:06into Slack,
00:20:07Salesforce,
00:20:08and Linear,
00:20:09you have to provision
00:20:10and manage access tokens.
00:20:13And most systems
00:20:14provision long-lasting
00:20:15credentials,
00:20:16meaning the agent
00:20:18effectively
00:20:18has permanent access.
00:20:21What's worse,
00:20:22those credentials
00:20:23are usually scoped
00:20:24to a human
00:20:25that created
00:20:26this access token,
00:20:27not the agents.
00:20:28So this means
00:20:29your agents
00:20:30have the ability
00:20:31to perform
00:20:32any action
00:20:33that you can,
00:20:34not the actions
00:20:35that they're designed for.
00:20:37That's why we built
00:20:38Vercel Connect
00:20:39to solve all
00:20:40of these problems.
00:20:42Vercel Connect
00:20:43gives your apps
00:20:44and agents
00:20:45secure,
00:20:47short-lived,
00:20:48minimally scoped
00:20:49access to your system.
00:20:51It's also integrated
00:20:52into Vercel's
00:20:54observability suite
00:20:55so that you can see
00:20:56how those tokens
00:20:57are being used.
00:20:59So let me show you
00:21:00how it works
00:21:01by setting up
00:21:01the first connectors
00:21:02for the agents
00:21:03I just described.
00:21:06Here,
00:21:07inside the Vercel dashboard,
00:21:09I'll start by creating
00:21:10the connectors
00:21:10that I want.
00:21:11I'll click
00:21:12on Create Connectors,
00:21:13and as you can see,
00:21:15we have a growing list
00:21:16of built-in connectors
00:21:17already available.
00:21:19It's including Slack,
00:21:21GitHub,
00:21:21Snowflake,
00:21:22Linear,
00:21:24Salesforce,
00:21:25and many more.
00:21:26And even if you don't
00:21:27see connectors
00:21:28that we don't yet support,
00:21:31you can create your own
00:21:32using OAuth
00:21:33and API keys.
00:21:35Let's start
00:21:35with Salesforce.
00:21:37I'll use the workspace
00:21:38that I want my agents
00:21:39to have access.
00:21:41And then hit Create
00:21:42Salesforce Connector.
00:21:46So what just happened
00:21:47looked simple,
00:21:48but behind the scene,
00:21:50Vercel is taking care
00:21:51of a lot of complexity
00:21:52on the integrations.
00:21:55So normally,
00:21:55connecting to Salesforce
00:21:56means you have to deal
00:21:58with OAuth workflow,
00:22:00token management,
00:22:01credential storage,
00:22:03and refresh logic,
00:22:04but with Connect,
00:22:05all of these
00:22:06are taken care of for you.
00:22:08This connection issues
00:22:09short-lived scope tokens
00:22:10on demand
00:22:11when the agents
00:22:12really need
00:22:12to perform action.
00:22:15So let's start
00:22:16creating the
00:22:17Linear Connector here.
00:22:19I'll go to
00:22:20Create Connector.
00:22:22I pick Linear.
00:22:23I make sure
00:22:24I have the right
00:22:25workspace.
00:22:27And
00:22:27what happened
00:22:30is that
00:22:31Vercel Connect
00:22:31enables me
00:22:32to choose
00:22:32exactly what permission
00:22:34I need to give
00:22:35to my agents.
00:22:35In this case,
00:22:37the agents
00:22:37need to comment
00:22:39and also create issues
00:22:40for the agents
00:22:41I described.
00:22:42So I'll only give it
00:22:43those permissions
00:22:44and then
00:22:45simply create
00:22:48the connector.
00:22:50So by restricting
00:22:51our agents
00:22:52to those permissions
00:22:53and scope upfront,
00:22:55we can make sure
00:22:56that this agent
00:22:57only performs
00:22:57the task
00:22:58that we want it to be
00:22:59and still have
00:23:01the powerful automation.
00:23:03So I'll finish
00:23:04creating the connector
00:23:05and I'll land it
00:23:06back in a details page.
00:23:08So you can see
00:23:08the project creation
00:23:09here.
00:23:10You can install.
00:23:11You can see
00:23:12the installation trigger
00:23:13and you can also see
00:23:14the usage data later.
00:23:16Let's quickly
00:23:17test our token here.
00:23:20So I'm going to test
00:23:22the user token.
00:23:24I'll click
00:23:25on test user token
00:23:26and I'll authorize
00:23:27my APIs
00:23:28and then I'll authorize
00:23:30on the linear side.
00:23:33And great,
00:23:33I'm giving back
00:23:34with a token
00:23:35that my agents
00:23:36needs to run
00:23:37the specific task.
00:23:38It has the expiry,
00:23:40it has the name
00:23:40and let's start,
00:23:44let's create
00:23:44the last connector
00:23:45that we want
00:23:46for our agents.
00:23:47I'll select Slack
00:23:48and I'll make sure
00:23:50that I have authorized
00:23:51workspace
00:23:52and I hit create.
00:23:54I'll install it later
00:23:56for a little bit
00:23:58I'll show you that
00:23:59and then going back
00:24:01to the page.
00:24:02So now that we have
00:24:03the three connectors
00:24:04installed,
00:24:05our agents
00:24:06now has access
00:24:07to everything it needs
00:24:08and it never stores
00:24:09long-lived credential.
00:24:12Back to you, Tomo.
00:24:16Thank you.
00:24:22With Vercel Connect,
00:24:23your agents can securely
00:24:24connect to your full range
00:24:26of internal systems
00:24:27including your CRM,
00:24:28ERP,
00:24:30HRIS,
00:24:31data warehouses,
00:24:33collaboration tools
00:24:34and more.
00:24:35But I want to double click
00:24:36on collaboration tools.
00:24:38This is how you interact
00:24:39with your agents
00:24:40and how your agents
00:24:41want to be able
00:24:41to interact with you
00:24:42and that's where
00:24:44the next primitive
00:24:44in the agent stack
00:24:45comes in
00:24:46which is chat SDK.
00:24:48The most powerful agents
00:24:50we've built at Vercel
00:24:51are effectively co-workers
00:24:52and we interact with them
00:24:54where work happens
00:24:55as it's happening.
00:24:57For many of you,
00:24:57that might be Microsoft Teams
00:24:59or Google Chat
00:25:00but for us,
00:25:01overwhelmingly,
00:25:02that's Slack.
00:25:03We believe that powerful
00:25:04collaboration tools
00:25:05like these
00:25:06are the primary way
00:25:07most humans
00:25:07will interact with agents
00:25:08but there's a problem.
00:25:10Even though these tools
00:25:11share a common set
00:25:12of core capabilities,
00:25:14they all have vastly
00:25:15different APIs
00:25:16and interfaces.
00:25:18Chat SDK provides
00:25:19an elegant abstraction layer
00:25:21enabling you to target
00:25:22dozens of apps.
00:25:23With just a single line
00:25:24of code,
00:25:25it enables your users
00:25:26to interact with agents
00:25:27across tools like
00:25:28Microsoft Teams,
00:25:30Google Chat,
00:25:31Slack,
00:25:32Discord,
00:25:33GitHub,
00:25:33Linear,
00:25:34Telegram,
00:25:35WhatsApp
00:25:35and so many more.
00:25:38NanoClaw helps companies
00:25:39run AI agents
00:25:40and they built their platform
00:25:42on the chat SDK.
00:25:43It's a single code base
00:25:44but it delivers agents
00:25:45across 15 different
00:25:47messaging apps.
00:25:49So that's the
00:25:50Vercel agent stack.
00:25:52It's a singular set
00:25:53of end-to-end capabilities
00:25:54covering everything
00:25:55needed to ship
00:25:56production agents.
00:25:58We've taken everything
00:25:59we've learned
00:25:59building agents
00:26:00over the past few years
00:26:01and turned those learnings
00:26:02into best-in-class primitives
00:26:04that work at Vercel scale.
00:26:06We love them
00:26:07and our customers do too.
00:26:09The stack is powerful
00:26:10and fills a real gap
00:26:12in the ecosystem
00:26:13and we believe
00:26:13every one of these primitives
00:26:14is best-in-class.
00:26:16But there's still
00:26:17a lot of complexity.
00:26:19Each of these primitives
00:26:20still needs to be wired up
00:26:21into a single cohesive agent.
00:26:24But what if they didn't?
00:26:26As we've built out
00:26:27the agent stack,
00:26:28this is the question
00:26:29we kept coming back to.
00:26:31At Vercel,
00:26:32we don't just want
00:26:33to build the world's
00:26:34most powerful agents
00:26:35and we don't want
00:26:36to just enable you
00:26:37to build the world's
00:26:37most powerful agents.
00:26:39we want building
00:26:40those agents
00:26:40to actually be delightful.
00:26:42And that means
00:26:43not just providing
00:26:44the world's most powerful
00:26:45agent stack primitives,
00:26:48but also the world's
00:26:50easiest and fastest way
00:26:51to stitch
00:26:52those primitives together.
00:26:54To make that possible,
00:26:56we built the newest member
00:26:57of the agent stack family.
00:27:00Welcome, Eve.
00:27:02Vercel's framework
00:27:03for building complete
00:27:04end-to-end production agents.
00:27:07Now, I have so much
00:27:08to say about Eve,
00:27:09but before I get into that,
00:27:10I want you to see it
00:27:11in action.
00:27:12So I'm excited to invite up
00:27:13Char, product lead
00:27:15for Eve,
00:27:15to show you how it works.
00:27:24Thanks, Tomo.
00:27:26Over the last year,
00:27:28we've taken everything
00:27:28we've learned
00:27:29about building agents
00:27:30and we packaged it
00:27:31into Eve.
00:27:32Let me show you
00:27:33how it works.
00:27:35I'll get started
00:27:35with a single command
00:27:37in my terminal.
00:27:38This command will scaffold
00:27:40an agent directory,
00:27:42install dependencies,
00:27:44and start an interactive
00:27:45chat session
00:27:46with the agent.
00:27:48Next,
00:27:49I'll configure
00:27:49our model provider.
00:27:51I'll use
00:27:52Vercel AI gateway,
00:27:55select a Vercel team,
00:27:59and link to a project
00:28:01in that team.
00:28:01And we're done.
00:28:05Now, let's test it.
00:28:07Who are you?
00:28:12Okay.
00:28:12That was fast.
00:28:14I just built
00:28:15a fully functional agent
00:28:17in less than a minute.
00:28:18Now,
00:28:23this agent runs
00:28:24with just two files.
00:28:26Agents.ts,
00:28:27which defines the model,
00:28:30and instructions.md
00:28:31that defines the agent's identity.
00:28:34And that simplicity
00:28:35is what makes it so easy
00:28:37to build agents with Eve.
00:28:39Now, let's build
00:28:40a real use case.
00:28:41Hedy just described
00:28:42a sales agent
00:28:43that processes call transcripts
00:28:45and updates Salesforce
00:28:46and linear.
00:28:47Let me show you
00:28:48how simple it is
00:28:49to build that agent
00:28:50with Eve.
00:28:52First,
00:28:53I'll give the agent
00:28:54an identity.
00:28:55In the instruction markdown file,
00:28:57I'll describe
00:28:58a go-to-market assistant
00:28:59that can create
00:29:00linear issues
00:29:01and update Salesforce
00:29:02opportunities
00:29:03based on call transcripts.
00:29:04And just like that,
00:29:10the agent has
00:29:11a new charter.
00:29:12Next,
00:29:13I'll give it tools
00:29:14it can use
00:29:14to take actions.
00:29:16The agent needs
00:29:16a linear tool
00:29:17so it can create issues
00:29:18and tools go
00:29:19inside the tools folder.
00:29:24And the file name
00:29:26is the name
00:29:27of the tool
00:29:27our agent sees.
00:29:34Let's define
00:29:35a linear tool.
00:29:37The description
00:29:37is what the agent reads
00:29:39to know when
00:29:39to use the tool,
00:29:41and the inputs
00:29:42are the data points
00:29:42that the agent
00:29:43has to fill in,
00:29:44like the title
00:29:45and the description
00:29:45of the issue.
00:29:47We can use
00:29:48the linear connector
00:29:50that Hedy set up earlier
00:29:52and then implement
00:29:53the linear API
00:29:54for creating an issue.
00:29:58I also want
00:29:59the agent to know
00:29:59how to create
00:30:00a linear issue,
00:30:01so I'll add a skill.
00:30:02And skills go
00:30:04to skills folder.
00:30:06This skill tells
00:30:07the agent
00:30:07that it should look
00:30:08for feature requests
00:30:09in the transcript
00:30:11and separate linear issues
00:30:12by topic.
00:30:16Great.
00:30:17Now, let's test the agent.
00:30:18It should be able
00:30:18to create linear issues.
00:30:25It's going to load
00:30:26the skill,
00:30:27create the issue.
00:30:28Great.
00:30:29The issue is created,
00:30:30and it should also
00:30:31show up in linear.
00:30:34Next, the agent needs
00:30:36to update Salesforce
00:30:37opportunities.
00:30:38We want to use
00:30:38Salesforce CLI
00:30:39to make updates,
00:30:41but that means
00:30:42our agent needs a shell,
00:30:43and we want the commands
00:30:44to be executed safely
00:30:46in isolation.
00:30:47So I'm going to configure
00:30:48a sandbox that the agent
00:30:50can use to make updates.
00:30:51This sets up a private VM
00:30:53for the agent to work in,
00:30:54walled off from everything else.
00:30:57I'll create a sandbox.ts.
00:31:00This is going to configure
00:31:01the sandbox,
00:31:02and then install Salesforce CLI
00:31:04in sandbox bootstrap,
00:31:07and also lock the doors.
00:31:09So the sandbox is only allowed
00:31:11to talk to Salesforce
00:31:11and nothing else.
00:31:15I'll give the agent
00:31:16the Salesforce tool
00:31:16just like we did with linear,
00:31:18but Salesforce updates
00:31:19need approval
00:31:20from someone on the team.
00:31:22So I'll add human in the loop,
00:31:24and this tool will always
00:31:26require the agent
00:31:26to ask for human approval.
00:31:32We'll use Hedy's Salesforce connector.
00:31:37And last,
00:31:38the tool execution function
00:31:39that will call Salesforce CLI
00:31:41and make updates
00:31:42to the opportunity record.
00:31:43Now, let's tell us
00:31:45the Salesforce step.
00:31:50I'll tell it to update
00:31:52Salesforce opportunity
00:31:53to closed one.
00:31:55That's human in the loop,
00:31:58and done.
00:31:59If you go back
00:32:00to Salesforce,
00:32:01refresh this page,
00:32:03it should say closed one.
00:32:07Awesome.
00:32:08Now, I need to make
00:32:09the agent available
00:32:10in Slack for the sales team.
00:32:12So I'll run slash channels.
00:32:15Select Slack.
00:32:18Yes, I want the Slack bot.
00:32:22This is going to install
00:32:23the Slack bot
00:32:24in the workspace.
00:32:28Deploy to Vercel.
00:32:34And done.
00:32:35Let's give it a test.
00:32:36I'm going to tag our agent.
00:32:41And that's it.
00:32:43We have a live,
00:32:45production-ready agent
00:32:46that our team can use
00:32:47in our collaboration layer of Slack.
00:32:50But we're not done yet.
00:32:51The agent is also fully observable.
00:32:54Let's take a look
00:32:55at the conversation
00:32:55we just had with it in Slack.
00:32:57So in the Vercel dashboard,
00:32:59I'll click in observability,
00:33:02agent runs,
00:33:05and I can see
00:33:06the entire conversation history.
00:33:09Token usage,
00:33:10length of the run,
00:33:11inputs, outputs,
00:33:13and the agent's reasoning.
00:33:14I can even drill down
00:33:16into each of the tool calls
00:33:17that the agent made.
00:33:19And that's it.
00:33:20We just feel
00:33:21the production-ready agent
00:33:22in five minutes.
00:33:24Back to you, Tomlin.
00:33:24thank you, Char.
00:33:33Eve truly feels like
00:33:34the missing Next.js
00:33:36for agents.
00:33:37It's opinionated
00:33:38based on everything
00:33:39we've learned
00:33:40over the past several years.
00:33:41But more importantly,
00:33:42it's open source
00:33:44and modular.
00:33:45It's built to work
00:33:46seamlessly with Vercel,
00:33:48with high cohesion
00:33:49to our infrastructure,
00:33:50but it's completely customizable
00:33:52so you can make it your own.
00:33:54You can swap out providers
00:33:55and keep the durable
00:33:56foundation underneath.
00:33:58Before Eve,
00:33:59every agent needed
00:34:00its own scaffolding
00:34:01and integrations.
00:34:02With Eve,
00:34:03every agent
00:34:04is just a directory
00:34:05laid out the way
00:34:06you already think
00:34:07about your code.
00:34:09Building an enterprise-grade agent
00:34:10used to take weeks.
00:34:11Now, as you just saw,
00:34:13it takes minutes.
00:34:15Eve is going to change
00:34:16the way many of us
00:34:18build agents.
00:34:18It's already completely
00:34:19changed the way
00:34:20that we build agents
00:34:21inside Vercel.
00:34:22But next up,
00:34:23I want to invite Malta
00:34:24to tell you about
00:34:25one of his favorite agents
00:34:27we've built
00:34:27that many of us
00:34:29are very excited about.
00:34:31Thank you.
00:34:31Thanks, Tomo.
00:34:42I'm Malta Vercel-Steel.
00:34:44The agent Tomo
00:34:45is most excited about
00:34:46is mine.
00:34:47We built it for Vercel first
00:34:49and now you can use it too.
00:34:51It's called Vercel Agent.
00:34:55Wait, wait, wait.
00:34:56Hold on.
00:34:57Shoot.
00:34:58I'm actually getting paged.
00:35:00Okay, wow.
00:35:01Okay, V0 is having
00:35:02a partial outage.
00:35:04Let me actually get
00:35:05to my laptop.
00:35:06Imagine I have a laptop
00:35:07and check this out.
00:35:09All right, so
00:35:10Vercel Agent already
00:35:12investigated this.
00:35:13Let's dive deeper.
00:35:15Okay, okay, okay.
00:35:16It found that the API
00:35:17key object is undefined
00:35:18at runtime.
00:35:19The errors appeared
00:35:20at the last deployment
00:35:21a few minutes ago.
00:35:23It recommends
00:35:24an instant rollback
00:35:25to the previous release.
00:35:26That's actually really smart
00:35:27because when in doubt,
00:35:28you always want
00:35:29an instant rollback
00:35:30as the right action.
00:35:31Let me approve this.
00:35:33Now, with my permission,
00:35:34Vercel Agent
00:35:35is rolling back
00:35:36the last production
00:35:37of deployment
00:35:37before the 500 errors
00:35:39showed up.
00:35:41All right.
00:35:41The rollback was successful.
00:35:43V0 is back online.
00:35:45Vercel Agent
00:35:45will now start working
00:35:46on a fix,
00:35:47but we don't have
00:35:48to wait for that.
00:35:49All right.
00:35:49Let's start over.
00:35:52Hey, my name is Malta.
00:35:53I'm Vercel's CTO.
00:35:54Vercel Agent is great
00:35:55because it autonomously
00:35:56detects incidents,
00:35:57only escalates
00:35:58when they are true positives
00:36:00and has the analysis
00:36:01already done
00:36:02by the time
00:36:02I get to my computer.
00:36:04Now, I do realize
00:36:06that the first question
00:36:08every CTO in this room
00:36:09will ask is,
00:36:11how is it safe
00:36:12to let an agent do this?
00:36:14It's the right question
00:36:16because most agents
00:36:17inherit the user's permissions.
00:36:20They run as you
00:36:21and they can do everything
00:36:23that you can do.
00:36:25A single bad prompt
00:36:27has your full blast radios,
00:36:29whether it comes from you,
00:36:31a teammate,
00:36:32or a confused sub-agent.
00:36:35Vercel Agent
00:36:35has a first-of-its-kind
00:36:37permission model
00:36:38that combines plan mode
00:36:40with granting permissions.
00:36:42Rather than asking you
00:36:43to approve actions
00:36:45one at a time,
00:36:46Vercel Agent plans
00:36:47what permissions
00:36:48will be required
00:36:49to complete a task
00:36:50and then ask you
00:36:52to approve them
00:36:52in a single coherent step.
00:36:55To roll back
00:36:56a broken deployment,
00:36:57Vercel Agent
00:36:58only asks for
00:36:58temporary permission
00:36:59to perform instant rollbacks,
00:37:01like we just saw.
00:37:02If it needs to purchase
00:37:03stale cash,
00:37:04it asks for temporary
00:37:04cash access
00:37:05to that project
00:37:06and nothing else.
00:37:08Every agent,
00:37:10every action
00:37:10runs in an isolated sandbox
00:37:11before it touches production.
00:37:14And anything
00:37:15that changes production state
00:37:16waits for a human
00:37:17to approve it.
00:37:19The agent never has
00:37:20more access
00:37:21than it needs
00:37:21for the task
00:37:22it's doing,
00:37:22but it also
00:37:23doesn't constantly
00:37:24prompt you
00:37:24for more permission.
00:37:26A great balance
00:37:27of progress
00:37:27and safety.
00:37:29It's the part
00:37:29I'm most proud of
00:37:30and also why
00:37:31I don't worry
00:37:32that Vercel Agent
00:37:33will be the cause
00:37:34on my next 2 a.m. page.
00:37:40Wait, wait.
00:37:41What's that going on?
00:37:42Someone not turning off
00:37:43their phone?
00:37:44It's 20.55.
00:37:45Oh my god.
00:37:46It was me.
00:37:46It was me.
00:37:48All right.
00:37:50Oh, I got a text
00:37:51from Guillermo.
00:37:51Dude, you perfectly
00:37:52know I'm on stage
00:37:53right now.
00:37:54Anyway, I realize,
00:37:55you all realize
00:37:56this is staged,
00:37:57but let me tell you,
00:37:58G texting me
00:37:59to complain about
00:38:00something not going
00:38:00right on our website
00:38:01is about as lifelike
00:38:03as things get.
00:38:05So anyway,
00:38:05let's copy-paste
00:38:06that message
00:38:07into Vercel Agent.
00:38:08Actually, let me add
00:38:10make no mistakes
00:38:11just for good measure.
00:38:15All right.
00:38:16Vercel Agent
00:38:17is now looking
00:38:17at speed insights
00:38:19for our website
00:38:19to see if anything changed.
00:38:22Oh, wow.
00:38:23Okay, there's a 4-second LCP.
00:38:25Something is majorly wrong.
00:38:26That's not good.
00:38:29Now Vercel Agent
00:38:29is going to review
00:38:31all the recent changes
00:38:32to narrow down the problem.
00:38:35And it found
00:38:36that we added
00:38:36in the wait
00:38:37on the Fabicon route.
00:38:40It checks performance
00:38:41on that endpoint
00:38:42and it confirms
00:38:44that an update
00:38:46caused the problem.
00:38:48It finished the investigation
00:38:49with the root cause
00:38:50and now this is
00:38:51and it looks like
00:38:51I'm going to have
00:38:52to talk with Matan.
00:38:54He's here in the audience
00:38:55somewhere.
00:38:56No worries.
00:38:57The good news is
00:38:58that Vercel Agent
00:38:59can write a fix
00:38:59in OpenAPR.
00:39:01When I approve it,
00:39:02the change will go
00:39:03into production.
00:39:05Okay, back to G.
00:39:06Fix the Chief.
00:39:07Zero lines of code
00:39:08written today
00:39:08by me.
00:39:11All right.
00:39:12Those are two examples
00:39:13of how I use
00:39:15Vercel Agent
00:39:16but can help you manage
00:39:17anything on the Vercel platform.
00:39:19When you ask
00:39:20to fix your build,
00:39:22it reads
00:39:22your deployment logs,
00:39:24finds the failing config,
00:39:25validates the fix
00:39:26in the sandbox
00:39:26and redeploys
00:39:27with your approval.
00:39:29Ask why your build
00:39:30got slower
00:39:30in the latest deployment
00:39:31and diffs the build
00:39:32against the previous one
00:39:33and tells you
00:39:34what changed.
00:39:36You can ask it
00:39:37to find the top
00:39:38accessibility issues
00:39:39in your project
00:39:39and it will run a review
00:39:40and open a PR.
00:39:43And if you ask it
00:39:44to fix the 500
00:39:45showing up in your logs,
00:39:46it will do the same.
00:39:48But what I love most
00:39:49about the agent
00:39:51is that you don't
00:39:52have to ask.
00:39:53Vercel Agent
00:39:54sees your app
00:39:54running a production
00:39:55so when traffic spikes
00:39:57or an alert fires,
00:39:58it will investigate
00:39:59immediately
00:39:59and bring answers
00:40:00to you
00:40:01just like it did
00:40:02for me
00:40:02when I got paged
00:40:03here on stage.
00:40:05The examples
00:40:06I gave you
00:40:07were normal
00:40:08infrastructure
00:40:08and performance
00:40:09problems.
00:40:10But soon,
00:40:10Vercel Agent
00:40:11will be able
00:40:11to call specialists.
00:40:13It will run
00:40:14DeepSec
00:40:14for deep security review
00:40:15across your whole
00:40:16code base
00:40:16or inspect
00:40:17your front end
00:40:18for design
00:40:19and UX quality.
00:40:21It's one agent
00:40:22that understands
00:40:22your stack
00:40:23and your infrastructure
00:40:24with exports
00:40:25it can call on demand.
00:40:27Vercel Agent
00:40:28is available
00:40:28in private beta
00:40:29today.
00:40:29Scan this QR code
00:40:30to request access.
00:40:33Uh-oh.
00:40:34Not again.
00:40:36Great.
00:40:37It's our CEO, Gene.
00:40:39All right.
00:40:42Okay.
00:40:42She's writing,
00:40:43I'm panicking
00:40:43a little bit.
00:40:44Did we actually
00:40:44flip the feature flags
00:40:45for everything
00:40:46we announced today?
00:40:48Now,
00:40:49obviously,
00:40:49I could not go
00:40:50and lock in the dashboard
00:40:51and look at Vercel
00:40:51flags myself,
00:40:52but what year
00:40:53is this?
00:40:532024?
00:40:55Let's ask Vercel Agent.
00:40:57What did we ship today?
00:40:59All right.
00:40:59We have
00:41:00enable Vercel services.
00:41:02Gene talked about this.
00:41:03Enable EVE.
00:41:03I'm so excited.
00:41:05I can't wait
00:41:05for all of you
00:41:06to try it.
00:41:07Enable Vercel Agent beta.
00:41:10You know,
00:41:10just talked about it.
00:41:11And then enable
00:41:11Vercel Passport.
00:41:13Actually,
00:41:14what is Vercel Passport?
00:41:16We haven't heard
00:41:17anyone talk about that yet.
00:41:19I might have
00:41:20just leaked
00:41:20the rest of the keynote.
00:41:21Anyway,
00:41:22to talk about
00:41:23Vercel Passport
00:41:23and other things,
00:41:24please welcome
00:41:25on stage, Gene.
00:41:26Thanks, Malta.
00:41:36Vercel Agent
00:41:37shows you
00:41:38what's possible
00:41:38when you build
00:41:39agents the right way.
00:41:41And every company
00:41:42in this room
00:41:43is going to build
00:41:43an agent
00:41:44just like it.
00:41:45There are two types
00:41:47of people
00:41:47in this room
00:41:48hearing that.
00:41:49Some of you
00:41:50are sitting here
00:41:51thinking,
00:41:52let's go.
00:41:53I am shipping
00:41:53that tonight.
00:41:55Then,
00:41:56there are the CIOs
00:41:57and CTOs
00:41:58in the room
00:41:58who are thinking,
00:41:59ugh,
00:42:00no.
00:42:01Because you can
00:42:02already feel
00:42:03what's coming.
00:42:04Shadow agents
00:42:05writing to systems
00:42:06with no audit trail,
00:42:08AI bot user
00:42:09closing tickets,
00:42:11spend you can't explain.
00:42:13Both of you
00:42:13are right.
00:42:15Building agents
00:42:16is easier
00:42:17than you think
00:42:17and way harder
00:42:20than you think.
00:42:21I'm going to tell you
00:42:22about that tension
00:42:23and what we've learned
00:42:25at Vercel
00:42:25by living it.
00:42:28Drew Bredvick,
00:42:29who works for me,
00:42:30is head of
00:42:30go-to-market
00:42:31engineering.
00:42:32In June 2025,
00:42:34he had the sexiest job
00:42:38in the world.
00:42:39His mandate,
00:42:40build the agents
00:42:41that transform
00:42:42how Vercel
00:42:43goes to market.
00:42:45It worked.
00:42:46A year later,
00:42:48agents are part
00:42:48of our daily workflow
00:42:49running at scale
00:42:51across our entire
00:42:52go-to-market organization.
00:42:54You did a great job,
00:42:55Drew.
00:42:56So today,
00:42:57I'm handing you
00:42:58a pager.
00:43:00Let me tell you why.
00:43:01What we learned
00:43:03was that agents
00:43:04are free.
00:43:05Free,
00:43:06as in free puppies.
00:43:08Everybody loves puppies,
00:43:10but they pee on your floor,
00:43:11they eat your furniture,
00:43:13and you can't go
00:43:14on vacation.
00:43:15Agents are free
00:43:16because anyone
00:43:17can prompt Claude,
00:43:19but agents are software,
00:43:21and we all know
00:43:22that software
00:43:22is never done.
00:43:24Someone has to maintain them,
00:43:26update models,
00:43:27and build new features.
00:43:29Building hundreds of agents
00:43:30taught us hard lessons.
00:43:33First,
00:43:34we saw the same problem
00:43:36solved over and over.
00:43:38Multiple agents
00:43:39needed to connect
00:43:40connect to the same
00:43:41internal systems.
00:43:42Each team built
00:43:43their own integrations
00:43:44from scratch.
00:43:46Second,
00:43:47each agent
00:43:48was reading
00:43:48from different knowledge bases.
00:43:50Our team would ask
00:43:51the same question
00:43:52and get different answers.
00:43:54Third,
00:43:55we had no visibility.
00:43:57No one knew
00:43:58how many agents existed,
00:43:59who built them,
00:44:00or what data
00:44:01they touched.
00:44:03Fourth,
00:44:04that lack of visibility
00:44:05also meant adoption chaos.
00:44:07I'm on Slack,
00:44:08I type at,
00:44:09500 agents pop up.
00:44:11I don't know
00:44:11what any of them do.
00:44:13And last,
00:44:13we also learned
00:44:14that chat
00:44:15isn't all you need.
00:44:17We started with the idea
00:44:18that Slack
00:44:19was a universal interface,
00:44:20and that was wrong.
00:44:22The agents
00:44:22that actually got used
00:44:24also had front ends
00:44:25for permissions,
00:44:26for visualizing data,
00:44:27for workflows,
00:44:28and for keeping humans
00:44:29in the loop.
00:44:31We learned
00:44:31that all of our agents
00:44:33had to work
00:44:33on day one
00:44:35and on day 100.
00:44:38And I'm happy
00:44:38to tell you
00:44:39that they do.
00:44:41We run over 100 agents
00:44:43in production
00:44:44at Vercel,
00:44:45and they're part
00:44:46of how we operate
00:44:47every single day.
00:44:49I want to tell you
00:44:51about 10
00:44:51of the most important ones.
00:44:53And the order matters.
00:44:55We started
00:44:55with the most obvious
00:44:56use cases
00:44:57and worked our way
00:44:58towards the agents
00:44:59that changed
00:45:00internal processes
00:45:01and transformed
00:45:02how we operate
00:45:04the go-to-market team.
00:45:07Vertex
00:45:07is our customer
00:45:08support agent.
00:45:10It resolves
00:45:11over 91%
00:45:13of Vercel's
00:45:14support tickets
00:45:15across the help center,
00:45:16Slack,
00:45:17and DocsChat.
00:45:19Deal1
00:45:20is our deal
00:45:21intelligence agent.
00:45:22It listens
00:45:23to every sales call,
00:45:25coaches our reps
00:45:25in Slack,
00:45:26and runs a post-mortem
00:45:27on every lost deal.
00:45:29The Deal1
00:45:30MCP has been called
00:45:3117,000 times
00:45:34this month.
00:45:36Draft Zero
00:45:36is our content agent.
00:45:38It writes
00:45:39the first draft
00:45:40of every blog,
00:45:41changelog,
00:45:42and customer story
00:45:43we publish.
00:45:45A Zero
00:45:46is our AEO agent.
00:45:48It tracks
00:45:48how Vercel's brand
00:45:50and content
00:45:50shows up
00:45:51across AI search.
00:45:53Every day,
00:45:54it runs hundreds
00:45:55of prompts
00:45:55across dozens
00:45:56of coding models.
00:45:59Revoa
00:45:59is our Salesforce
00:46:00update agent.
00:46:02It pushes
00:46:02critical record
00:46:03changes
00:46:03into Salesforce
00:46:05with a human
00:46:05in a loop.
00:46:06It saves nine hours
00:46:08of time
00:46:08for our RevOps team
00:46:10every day.
00:46:12Penny
00:46:12is our finance
00:46:14and ops agent.
00:46:15It has access
00:46:15to our billing platform,
00:46:17payment provider,
00:46:18and monitoring systems.
00:46:20It saves our finance
00:46:21and on-call
00:46:21engineering team
00:46:22hours answering
00:46:23billing tickets.
00:46:26And next,
00:46:27there's V.
00:46:28V is our routing agent.
00:46:30It routes requests
00:46:31to all of the
00:46:32other internal agents.
00:46:34Remember that
00:46:35adoption problem?
00:46:36We still have
00:46:37100 agents,
00:46:38but V
00:46:39is the front door
00:46:40to all of them.
00:46:41Ask V a question,
00:46:43and it picks
00:46:43the right agent
00:46:44for the job.
00:46:46The last thing
00:46:47I'm going to show you
00:46:47are the ones
00:46:48I want to spend
00:46:49real time on
00:46:50because I think
00:46:51every single one
00:46:52of you is going
00:46:52to want to build
00:46:53one of them.
00:46:55First,
00:46:56DZero,
00:46:57our data analyst.
00:46:59DZero gives
00:46:59our entire company
00:47:0124-7 on-demand
00:47:03at analytics
00:47:04and data science.
00:47:05Anyone at Vercel,
00:47:07engineers,
00:47:07AEs,
00:47:08finance,
00:47:09support,
00:47:10can run analytics
00:47:11on our data
00:47:12from our warehouse
00:47:13without filing a ticket
00:47:14or waiting on the data team.
00:47:17Users can ask
00:47:17simple questions
00:47:18like how many leads
00:47:19we got from a campaign,
00:47:21and DZero writes
00:47:22and runs basic SQL.
00:47:25But DZero is also
00:47:26a data scientist.
00:47:28If users need
00:47:28statistical analysis done,
00:47:30it spins up a sandbox
00:47:31and runs Python
00:47:33to generate reports.
00:47:35DZero is the most used
00:47:37internal tool
00:47:38at Vercel.
00:47:39It answers 30,000
00:47:41questions a month.
00:47:43And it's safe at scale.
00:47:45DZero doesn't run
00:47:46in God mode.
00:47:47Every query is scoped
00:47:48to the user's permissions.
00:47:51If you can't see a table
00:47:52in Snowflake,
00:47:53DZero can't show it to you
00:47:54either.
00:47:55Under the hood,
00:47:56DZero needed a semantic layer.
00:47:59And even though
00:47:59you can ask questions
00:48:01in Slack,
00:48:02it needed a UI
00:48:02so people could explore charts
00:48:04and drill into data.
00:48:06Next is Athena,
00:48:09our sales cockpit.
00:48:11Salesforce announced headless.
00:48:13We've been running on it
00:48:14for months.
00:48:15Athena picks accounts,
00:48:17plans outreach,
00:48:18tracks signals,
00:48:19and runs the weekly motion
00:48:20for every AE at Vercel.
00:48:23Shortly after it went live,
00:48:25pipeline nearly doubled.
00:48:27Every AE uses it
00:48:29every single day.
00:48:31Under the hood,
00:48:32Athena needed
00:48:33the same semantic layer
00:48:34that DZero needed,
00:48:36plus durable workflows
00:48:37and secure connections,
00:48:39plus a UI,
00:48:40because GTM agents
00:48:41are more effective
00:48:42with pixels and buttons
00:48:44than just a Slack channel.
00:48:47And last,
00:48:48Lead Agent,
00:48:49our autonomous SDR.
00:48:51We trained Lead Agent
00:48:52on the playbook
00:48:53of our best SDR,
00:48:55and now it runs
00:48:56the playbook itself
00:48:5724-7.
00:48:58You may have seen
00:48:59the headline last year.
00:49:01We took 10 SDRs
00:49:02down to one.
00:49:04That was Lead Agent.
00:49:06We redeployed
00:49:07those nine reps
00:49:08into bigger roles
00:49:09and raised our quota.
00:49:10We've seen a 32X ROI,
00:49:13and it costs $5,000
00:49:15a year to run.
00:49:16It performs
00:49:17at the 90th percentile
00:49:19of our reps,
00:49:20and one engineer
00:49:21maintains it part-time.
00:49:23Under the hood,
00:49:25Lead Agent runs
00:49:26the stack Tomo
00:49:27just showed you.
00:49:28AI SDK,
00:49:29Workflow SDK,
00:49:30and Chat SDK
00:49:31for Slack.
00:49:32And it's open source,
00:49:34so you can go build
00:49:35your own today.
00:49:37Building these agents
00:49:38was easy
00:49:39because we used Next.js
00:49:40and our agent framework,
00:49:43Eve,
00:49:43and running them
00:49:44was never a problem
00:49:45because they run on Vercel.
00:49:47Remember when I said
00:49:48building agents
00:49:49was harder
00:49:50than you think?
00:49:52The hard part
00:49:53is everything around them,
00:49:55who can access them,
00:49:57how they authenticate,
00:49:58what data they can touch,
00:50:00and providing all of it
00:50:01to your security team.
00:50:03So we spent the last year
00:50:04building the platform
00:50:05that makes that easy,
00:50:07and today,
00:50:08we're making it available
00:50:09to you.
00:50:11I'm excited to announce
00:50:12Vercel for enterprise apps
00:50:15and agents.
00:50:17It's the Vercel developer
00:50:19experience you love
00:50:20for everyone at your company
00:50:22with identity
00:50:23and access built in
00:50:24and the option
00:50:26to run it
00:50:26in your own AWS tenant.
00:50:29Like all of our products,
00:50:30it's portable
00:50:31because it's framework
00:50:32and model agnostic,
00:50:34and we built this
00:50:35for ourselves first,
00:50:36and now it's a platform
00:50:37you can build on too.
00:50:39Let me show you
00:50:40the three most important pieces.
00:50:43Before today,
00:50:44you had to manually provision
00:50:45and off-board
00:50:46every Vercel user.
00:50:48Enterprise-managed users
00:50:50gives you central control
00:50:51over employee access
00:50:52to Vercel and VZero
00:50:54with a full audit trail
00:50:56of every action they take.
00:50:59I said earlier
00:51:00that anyone can prompt Claude.
00:51:02The reality is
00:51:03your employees
00:51:04are already doing this,
00:51:05whether you know it or not.
00:51:07That's called shadow IT,
00:51:09and AI has already caused
00:51:11major data breaches
00:51:12in the enterprise.
00:51:13So even if you can control
00:51:15who builds with AI,
00:51:16you still have to limit
00:51:18the access to apps
00:51:19and agents they build.
00:51:21That's why we built
00:51:22Vercel Passport.
00:51:24It puts every internal app
00:51:26and agent behind your IDP
00:51:28by default.
00:51:29Internal tools
00:51:30stay internal.
00:51:32Employees can only see
00:51:33the apps they need,
00:51:35and nothing is publicly exposed.
00:51:38And you can deploy
00:51:39all of this
00:51:40to your own AWS tenant.
00:51:42You heard that right.
00:51:43You can even run
00:51:44Vercel functions
00:51:45in your own AWS account.
00:51:47Whatever you build
00:51:48with agents
00:51:49stays inside
00:51:50your security boundary.
00:51:53The teams
00:51:54who ship fast,
00:51:56securely at scale
00:51:57are the ones
00:51:58who will win.
00:51:59We built enterprise apps
00:52:01and agents
00:52:01so that can be you.
00:52:03G,
00:52:04back to you
00:52:04to wrap up.
00:52:13Thanks, Gene.
00:52:14Thanks, Gene.
00:52:15Killed it.
00:52:17Today,
00:52:18we showed you
00:52:18that Vercel
00:52:19is a platform
00:52:19where you can build
00:52:21and run everything.
00:52:23Any kind of website,
00:52:25app,
00:52:26or agent.
00:52:28Your backend frameworks
00:52:29can run at scale.
00:52:31With Vercel services,
00:52:33you can develop,
00:52:34preview,
00:52:35and deploy
00:52:36backends
00:52:36and frontends
00:52:37together.
00:52:40With EVE,
00:52:41anyone on your team
00:52:42can build an agent
00:52:43in minutes
00:52:44or seconds.
00:52:47Vercel Connect
00:52:48gives those agents
00:52:49secure access
00:52:50to all of your services.
00:52:54Vercel Passport
00:52:55makes sure
00:52:55internal agents
00:52:56and apps
00:52:57stay internal
00:52:58behind your IDP.
00:53:01and you can run
00:53:02all of this
00:53:03in your own
00:53:04AWS account.
00:53:07You can start today.
00:53:09Ask your coding agent
00:53:10to install
00:53:11the Vercel plugin.
00:53:13Then you can build anything.
00:53:16A website,
00:53:17an app,
00:53:18an agent,
00:53:18and ship it
00:53:19at a global scale.
00:53:21And Vercel Agent
00:53:22will keep an eye
00:53:23on production.
00:53:25This is
00:53:26agentic infrastructure.
00:53:28We can't wait
00:53:29to see
00:53:30what you ship next.
00:53:31Thank you.

Key Takeaway

Vercel is shifting its infrastructure to an agentic-first model, providing a stack of specialized primitives—including AI Gateway, Workflow SDK, and the Eve framework—that allow developers to build, deploy, and securely manage autonomous agents at scale.

Highlights

  • Agentic workloads on Vercel have doubled since the beginning of 2026, with AI Gateway volume growing from 2 trillion to 20 trillion tokens per month.

  • Teams using coding agents deploy to Vercel five times more frequently than teams that do not.

  • The AI SDK has surpassed 15 million weekly downloads and serves as the universal toolkit for model-agnostic agent development.

  • Vercel Sandbox provides isolated, production-grade micro-VM environments to securely execute untrusted code generated by AI.

  • Vercel Connect enables secure, minimally scoped access to internal systems like Salesforce and Linear without storing long-lived credentials.

  • Eve is an open-source framework designed to scaffold and manage end-to-end production agents in minutes rather than weeks.

Timeline

Rise of Agentic Infrastructure

  • Agentic workloads on the platform have doubled in 2026.
  • Deployment volume from coding agents has increased 17 times over the last six months.
  • Vercel Services now allows unified development and deployment of full-stack applications with a single command, vcdev.

The evolution of the web from static sites to dynamic applications has culminated in a new era of software that can think autonomously. Vercel is building infrastructure specifically for these agentic workloads, which now account for over half of all platform deployments. The new Vercel Services feature resolves the complexity of wiring frontends and backends together by offering integrated local development and full-application preview URLs for every commit.

The Agent Stack Primitives

  • AI Gateway functions as a 'token delivery network,' offering routing, observability, and cost management across 35 different models.
  • Workflow SDK provides durable compute that automatically manages retries, state persistence, and long-running asynchronous tasks.
  • Vercel Sandbox enables secure execution of untrusted AI-generated code within isolated micro-VMs.
  • Vercel Connect facilitates secure, short-lived access to external tools like Salesforce and Slack without managing permanent credentials.

Modern software no longer follows a simple request-response model; it requires persistent agents that interact with various systems. Vercel built a stack to address these specific needs: an AI Gateway to optimize token delivery, a Workflow SDK for reliable long-duration tasks, and sandboxing for safe code execution. These tools allow complex applications, such as those used by Brex or Notion, to operate securely in production environments.

Eve and Vercel Agent

  • Eve is an open-source framework for building, testing, and shipping production-ready agents.
  • Vercel Agent is an autonomous system that detects production incidents and performs automated rollbacks with human approval.
  • The Vercel Agent uses a unique permission model that combines 'plan mode' with granular, temporary token granting.

Eve simplifies agent creation by standardizing project structure, allowing developers to scaffold functional agents in under a minute. Complementing this, the Vercel Agent platform manages live infrastructure, using observability data to investigate outages and suggest fixes autonomously. This combination provides a balance of high-velocity development and strictly controlled security boundaries.

Enterprise Governance and Scale

  • Vercel for Enterprise Apps and Agents provides centralized identity and access control for all internal tools.
  • Vercel Passport secures internal agents by placing them behind an IDP by default.
  • Platforms can now be deployed within a customer's own AWS tenant to maintain strict security boundaries.

Managing agents at scale introduces significant security risks, including shadow IT and unmonitored system access. Vercel's enterprise offerings address these challenges by providing full audit trails, mandatory identity provider integration, and the ability to run Vercel functions within private cloud environments. This ensures that even the most autonomous agents remain within organizational security compliance standards.

Community Posts

View all posts