Summary
vc dev command; front-end and back-end develop together locally, with preview URLs for entire stacks including backend-only componentsagents.ts (model/instructions) and instructions.md (identity/charter)Notes
Transcript
Stay hydrated. Talk to each other, learn from each other, and have a great time.
The venue that we chose has a lot of significance for us and Chris is going to talk about it in more detail but it is both an emotional connection and a physical connection because I don't think we could have picked a venue that would have been more on brand than this one. So Chris, what's up? What happened here? First and foremost, hello and welcome. I am Chris Williams, I am the head of our forward deployed engineering team at Vercel and Ich lebe Bern.
Yes, I have to tell you, it is an amazing city. I've been coming here since 2000, oh actually 2006. And I've been, I have to say, there are many origin stories that have happened in Berlin, but also literally in these halls. And so I'm going to give you a little bit of a background about it. You are in the Old Mint, "Alze Münze," as part of the old coin. And when it was first opened up, they would hammer the coins into existence.
As time went on, and as we are experiencing right now, there are great revolutions and changes in how that process was done, leading to full automation, which we are experiencing in the agentic revolution. Also, here about 16, 15 years ago, I had the great privilege to launch the JS logo. So JavaScript itself got its identity in Berlin. Not many people know that. This is the birthplace of the JavaScript logo and the identity that you see the yellow JS thing.
Also, And partner to our wonderful company, the origin story of Vercel has its roots here in Berlin as well. Malte Ubel, who is our CTO, he'll be up here shortly, he ran into a young gentleman at that point in time who was working on WebSockets and presented on the stage that Malte and two other gentlemen had put together in JSConf EU. That started a tremendous friendship That individual is Guillermo Rauch, who is our CEO.
And the two of them have gone on to do absolutely wonderful things. It is with great pride and pleasure that I get to introduce my great friend, our CTO, Malte Obel to the stage. Hey everyone, so great to be here. Can I get the prompt to get out of the big screen? By the way, hence the good comments, I'm originally from Germany, but it's good to be here. All right, so since SHIFT last year, The world has changed.
We are writing less skill by hand, but more ideas are coming to life than ever before. Six months ago, less than 3% of Versal's deployments were triggered by coding agents. Since then, that number has grown 17x. Over half of diplomas from Rassel-Nadl come from agents. But the bigger shift is With those deploying, HSR deploying, Since the beginning of the era, agentic work with Roussel doubled. Volume of the ad gateway grew from 2 trillion to 20 trillion tokens per month. The shift we're seeing is that we're using agents to deploy software that can think.
We're shipping agentic sites and apps. We use an aegis to ship ages. This is a profound change. I want to go back in time and show you how we got there. Our new world is ingentive. But it all started with websites. I remember building simple websites with HTML, CSS, and JavaScript as a kid. I was so excited to put my creations online That is what the web is all about. Sharing our ideas with the world. We've got the first version of Vercel.
to make that easy for anyone. It was infrastructure for pages, caching, and global content delivery. Today we serve some of the most innovative companies in Europe. Vekoni, Astra AI, and SAP. But the world's dynamic and as the web proved, science-oriented data-powered applications. So, we're selling infrastructure for servers, APIs, and databases. Now to sum up, EF and Tools for Humanity were an apps and scale on Perseverance.
For over a decade, we've been building our style into a cloud where you can run everything. Our vision is to build a true full stack platform and Last year, we shifted some of the huge steps towards that vision. We brought in some of the best Python developers in the world to take a look at our core infrastructure for backend frameworks. You can now run backends like FastAPI, Flask, ExpressJS, and HONOR at scale on your cell.
We build on a complete agentic layer. You can now run longer duration functions on fluid compute Make workflows durable and spin-off secure sandboxes. I just lost my boxes here.
Guys? All right, there we go. Oh yeah, that's good. So, for making workbooks doable, ship spay/not-secure sandboxes and host things, M60 servers, and make your app available to any agent. And we added the best database in the world to our platform. You can manage Amazon Aurora, Aurora Day SQL, and DynoDB, along with hundreds of high-quality integrations like Oversearch, all over the site. This means you can run any architecture you want on our infrastructure.
OpenAI, a company you might not be familiar with, runs the next-gen S/4HANA with a Python background yourself, delivering chat-to-BT to hundreds of millions of users every day. So in addition to front ends, You can host backend-only services on your cell, like REST APIs written in typescript or Python. You can deploy workflows that handle long-running, asynchronous work, even with a human in the loop.
And you can host huge step-by-step messages per month. But you don't always tell us It's still too hard to wire all these pieces together. And so we fixed it. Today I'm excited to announce for solution Services. It's a developed experience that helps you Oh my god, guys. There you go. It's a development experience that you know and love for yourself for your full stack applications. You can now develop your front-end and back-end together with one command, vcdef.
Everything spins up locally. When you push a commit, you get a preview URL for your entire round. Not just the front end, even backend-only components generate full preview you can test before your ship.
And all of the services you can deploy can talk to each other privately without ever touching the private internet.
You can run all of your microservices on Vercel and everything just works.
But in the ancient era, these services are becoming autonomous. Websites and apps used to respond to user input with logic. But now they have agents inside that can understand intent and take autonomous action. Every new generation of software demands a new generation of infrastructure. And the age of the era is no different. agentic software is agentic infrastructure. That's exactly what we're building for some. The genetic infrastructure has three parts.
first vercel is Recording Agent's deployed software. When you ask plot code or codex where to apply, you get "Restall." Because we're self-built for the way agents work. Second. Resell is where you build and deploy your agents. We give you every tool you need to build and run apps and agents in production, securely, and at scale. One platform for everything. And third, Resel itself is automated by agents.
Resel runs through apps and production, and traffic and traces, observability and anomalies. That data gives our agents the context they need to investigate autonomously and then surface pull requests, not just alerts. Teams using cloud code deployed themselves five times more frequently than teams who don't. That's because we set the standard for developer experience and now we're doing the same for agents that is developed by using.
Coding shoes love the point of herself because we give them everything they need. When you ask your agent to verify its work, it needs a live URL to test. Roussel gives every deployment a dedicated, secure, free view environment. When you ask an agent to ship experiments, they need to roll them out safely. Resel gives every experiment a feature flag and gives you control and confidence with instant rollback.
And you don't want your agent wasting time clicking around in the dashboard. An agent is most efficient when every part of the platform is available in its own language. That's why Masala gives it. That's what Roussel gives us through our CLI, API, and MCP. This year, an engineer at Meta opened CloudCo SU and asked us to build an internal tool. When it was time to test it, he asked Claude where to deploy.
and a push to herself. A week later, everyone on the team was deploying to RESTELL. Within a month, RESTELL was the go-to platform for all of Meta's super-intelligent slaps. Meta's frontier AI product was born on their cell. That happened even though Meta had already built their own platform for decades. But even the most powerful infrastructure doesn't natively speak the language of agent. or sell ducks.
Injective infrastructure is the platform we build But what matters are the products you fill the top with. For now, I want to invite Suyak Rao to the stage to show you how. Thank you very much.
Hi everyone, it's great to be back here in Berlin. Madhush has talked about why coding agents choose to deploy onto our agenting infrastructure.
AISDK was released three years ago. And it has been in active development since then. Did you know that Berlin was the birthplace for AISDK? Shout out to Lars Grammel, who is the creator of AISDK and one of our critical contributors to our library. Lars. Thank you. If you have questions,Harsh is here to answer all those. Ah. Okay, AI SDK is the universal toolkit for building AI frameworks, apps, and agents.
It's platform, model and framework agnostic. It allows you to build text, images, speech, video, and more. But it also makes it super trivial to add things like streaming and fallbacks to a project. The AISDK has been downloaded over 16 million times every week. And it has become the standard way to access any model From any provider in both TypeScript and Now, the Python has grown. It's used at massive scale across all over the world In companies like Rex, Rex!
All without needing to change your bio. You get to focus on building your features and AI SDK handles the rest. And today, AI SDK takes its biggest step here. I'm super excited to announce AI SDK 7 right here. The first models just returned text, but today they can reason, call tools, and work across files, sandboxes, and many hunters. AIC case gives you provider level abstraction for the entire life cycle.
You can build AI apps, agents, and even entire framework No lock-in at all to any single lap. Reasoning control lets you set reasoning effort once across parameters. Tools give you type tool context and human approvals. HARDEST support means that you can drop in plot code, codex, and pi into your AISDK. Durability and sandboxes make your agents long running and secure. Telemetry traces every step, every call, and every run.
And so the CDN was won. CDN Suite became the internet's performance and reliability platform. Intelligent Rallying Distributed edges. Failover, observability, centralized policies, and so much more capabilities that it brought to us. And today, AI use cases have outgrown the idea that every token should be fetched from the modelers. Tokens have become a production dependency and monolabs They've become the de-original.
In fact, our AI Gateway Production Index has shown us that team-striping agents at scale use more than 35 models Impression.
This means that routing isn't just a feature, it's actually how you run AI in production. Our customers, AKQA, with the chatGPT app for Starbucks using Next.js and AISDK, they are getting it.
It's an interactive bot that helps people discover new drinks that they may like. They use different models for different tasks.
Optimizing for cost And all. Gemini Flash, handles retrieval and summarization. Then recommendations are sold by CP5 And even more models are used to generate resonance. The best part? All with a single AI gateway in the app.
The African way today sells more than a trillion tokens every day.
Timeouts, drop connection, last day. But without a primitive, for dear believing, you're gonna spend a lot of time hand stitching things like V-drys and state persistent, it can get really ugly and cumbersome. That's exactly why we built the World's Greatest Day. We like to say it provides infinite computer ability. been able to build long running apps that R-90 Suspend resumed retry in the state of East.
DoorDash, our customer, uses WordpressDK to run traditional AEO jobs so that their data in the apps are up to date.
And our customer Flora built an entire AI design platform on workflow SDK. The platform helps designers generate visual images I assume Agents sign up across multiple providers, sometimes 50.
That's where our next instinctively comes in. We're still silence. Elements have become really good at generating code. But they're still untrusted. You don't want to run untrusted code in the same environment that has access to your production system. We need a special production grid environment Something that is isolated, secure, And Warsaw is no stranger to this problem of Unrestricted code execution at scale because of our review department's events.
The same MyCellular computer environment today is available to you and to your agents. Resultant Box is built on top of fluid compute. It supports dynamic, real-time workloads for agents.
It's already being used at scale by some of the largest companies in the world, including Notion.
But the code has to run safely. That's why Notion runs custom agents on Wurzel sandboxes. Each agent gets his own general purpose computer.
And everything outside of that is fully secure. So now our agents can execute code securelyInside time running workflows, But in order to do anything useful, they have to connect to data and tools. This is where the next layer of our agent stack comes in. We're Cell Connect. Well, Self-Connect is a brand new building block that allows Your agent to securely connect to all the data and systems you need access to. But it is secure access.
But here's the most important part. Instead of long living secrets, Every connection uses sharp Tom? Narrowly scope tokens, security best practices.
An agent never has standing access to your systems and never gets the data that it should.
With Vercel Connect, your agents can securely connect to a full range of internal systems, data warehouses, and collaboration tools. Now, let's double click on the collaboration rules aspect. Which is hollow. You interact with your agents and how your agents interact with you. So the next up in our agents pack primitive is SaaS. That is the key.
Chalice2K provides a very elegant abstraction layer for enabling dozens of apps with just a single line of code. R-R. Our customer Dan Appal helps companies run AI agents and they build These platforms are javascript Okay. So this is Agent, stop. We provide lots of different primitives for end-to-end capability for building production agents. Our customers love it and so do we too. Um... So one of the things that we've been thinking of As we build these primitives is that you start You still have to stitch all of these things together.
Thank you so much, Suyog. Over the last year, we've taken everything we've learned about building agents and packaged it into EVE. Let me show you how it works. I'll get started with a single command in my terminal. This command will scaffold an agent directory, install dependencies, and start an interactive chat session with the agent.
Wow, that was fast! I just built a fully functional agent in less than a minute! This agent runs with just two files. Agents.ts configures the agent's model and instructions.nd defines the agent's identity. And that simplicity is what makes it so easy to build an agent with Eve. Now let's build a real use case. Let me show you how simple it is to build that agent with Eve. First, I'll give the agent an identity.
In the instructions.md file, I'll describe a go-to-market assistant that can create linear issues and update Salesforce opportunities based on call transcripts. And just like that, The agent has an updated charter. Next, let's give it tools it can use to take action. The agent needs a linear tool so it can create issues. and tools Go inside the Tools folder. And the filing is the name, the tool, our agencies.
Great! The issue was created. Next, the agent needs to update Salesforce opportunities. We want to use the Salesforce CLI to make updates, but that means our agent needs a shell and we want the commands to be executed safely in isolation. I'm going to configure a sandbox the agent can use to make updates.
So I'll add a human in the loop. This tool will always require the agent to ask for human approval. We'll use the Salesforce connector and last the tool execution function that will call Salesforce through the CLI and then update this opportunity record. Now let's test the Salesforce connector. I'll ask the agent to update the Salesforce opportunity.
And that's it. We have a live production ready agent that our team can use. But we're not done yet. The agent is also fully observable. Let's take a look at the conversation we just had with them. In the Vercel dashboard, I click on observability. And then Agent runs. And here's the conversation. I can see the entire conversation history Token usage, the length of the run, input, output, and agent reasoning.
I can even drill down into each tool call the agent made. And that's Eve. We just built a production agent in five minutes. Back to you, Sea Oak.
Thanks, Malavika. ETH is NetJS for agents. It's opinionated based on everything we've learned Building agents in the past two years.
But completely customizable to make it your own. You can swap providers and keep the durable foundation underneath.
Smile. We bill it for ourselves first, but now you can use it to It's called Rassel agent. Hold on, hold on, sorry. It's not going well today, right? I just got paged. OK. OK. Wow. OK. v0 is having a partial outage. I'm going to go to my very virtual laptop and check this out. Hi, hi. I see. So, Grisel is already investigating this issue.
It found that the ADI key object is undefined at runtime. The error appeared after the last appointment a few minutes ago. And it recommends an instant rollback. And when in doubt, it's a robot, it's always the right thing to do. So let me improve that.
So with my permission, Purcell agents rolling back for the last production deployment before the 500 era showed up. Alright, Roblox was successful, we searched back in line, we saw a huge no-start, to work on the fix, and then we started working. Cool. The salvage is great because the "autonomously attacks incident" only escalates with their true positives, and has the analysis already done by the time I get to my computer.
Now understand, the first question every CEO has in the room to ask is, "How is it safe to let an agent do that?" It's the right question because most agents inherit the user's permissions. They run as you and can do everything you can do. A single bat prong has your full blast radius. Whether it comes from you, a teammate, or a confused sub-agent. Vercel Agent has a first of its kind permission model that combines plan mode with granting permissions.
Rather than asking you to approve actions one at a time, RecyAce's plans for permission are required to complete the task and ask you to approve them as a single, coherent step. To roll back a broken deployment, restart agent after temporary permission to perform instant rollbacks. If it needs to purchase stale cash, it asks for temporary cash access to that project and nothing else. Every action requires an isolated sandbox before it touches production.
I see, okay, there's a four second LCP.
It checks performance on that endpoint. and confirm the update caused by the problem. It finished the FAS investigation with a root cause analysis and looks like I'm going to have to talk to my colleague, Martin. And again, the good news is that we're still agents in pride, so fix and open it, y'all. But after a bit, the change will go into production. I'm going back to my past state, fix the cheese. We're zeroing in on the ball-free buggy today.
But soon, Percell agent will be able to run specialists. They will run, for example, deep-set, where deep-set should be reviewed across your whole codebase, or inspect your frontend for design and UX quality. It's one agent that understands your stack, your infrastructure with experts that can call on demand. Roussel Agent is available in private beta today. Scan this QR code to request agents. Ah, no text message. Keeps happening. Okay, this is Jean, our CEO.
I worked. Today, a year later, agents run every aspect of our go-to-market value chain. Every facet, every feature, every workflow. Congratulations Drew, congratulations to Voto Market Engineering and for all of those efforts, your award is... We're going to give you Peter. Let me tell you what. What we've learned, and a joke we have internally, is that agents are free. Free as in free puppies. Everyone loves puppies, puppies are great, but puppies can't really afford it.
Puppies eat furniture. Puppies make it so you can't go on vacation. Agents are free because anyone can work hard. Anyone can build agents. But agents are software, and software, like puppies, need love, care, and feed. Someone has to maintain that software. Someone has to update those models. Someone has to build new features. Now after building hundreds of agents at Roussel, we learned some hard lessons.
First, Multiple agents need to connect to the same internal systems. And each team that was building these agents had to rebuild those integrations over and over again from scratch. Second, each agent was reading from different knowledge bases, different stores of context. So two different agents from two different teams could be asking the exact same question and give different responses based on the context they have.
Three, problematically, we have no visibility or limited visibility. No one knew how many agents existed, no one knew who built them, what data they touched, what they could do. Four, relatedly, that lack of visibility also meant adoption chaos because we have all of this agent sprawl and you could go to Slack, type in@ and see 500 agents and not know what to do or which agent to use when. And finally, we also learned that chat is not all unique. Now when we started this, our vision was that Slack would be the universal interface, that Slack would be the home for all these agents.
And that got us very far, but it doesn't go far enough because the agents that actually got used at Vercel, the ones that got used the most at Vercel, had front ends. Dynamic, rich, authenticated, curated front ends for people to visualize data, manage human in the loop, manage these complex workloads. We haven't solved these problems. Because we knew that our agents had to work on day one. And on day 100.
And today, I'm happy to say that at Roussel, they do. Today we run over 100 agents in production at Versace. and they are central to how we operate every single day. So what I want to do is I want to talk about 10 of those agents, 10 of the most important agents. And the order here matters because we started with the most obvious use cases that people are going to reach for. And then we worked our way towards the long tail, towards agents that transformed every nook and cranny of the internal process and how we operate as a company.
Starting with... Starting with So Vertex is our customer support agent. Vertex handles 91% of the results for tickets. It works in Slack, it works in the Help Center, and it works on Docs Chat.
Deal One. Deal One is our deal intelligence agent. It listens to every sales call, it coaches our reps in Slack, and it runs a postmortem on every single loss deal. The Deal One MCP has been called 17,000 times in the past month. The graph zero Our content agent. It writes the first draft of every blog post, every change log, every customer story we publish, every external comms, including the comms that went out to this very conference.
Ae0 is our AEO agent. It tracks how Cursale's brand and their content show up across the iSearch. Every day, Ae0 funds hundreds of prompts across dozens of coding agents and gives us feedback on where we can improve.
Revoa is our Salesforce update agent. It pushes critical record changes into Salesforce for the human in the loop and saves nine hours of time each day for our love team.
Penny is a our finance and ops agents. It has access to our billing platform, our payment fiber, our monitoring systems, and saves our on-call engineering teams and our finance teams hours a day answering billing questions. And most importantly, one we're particularly excited about these days is... V V is our routing agent. It routes requests to all of our other agents. Because what I talked about earlier, the adoption problem still applies.
DZero gives our entire company 24/7 on-demand analytics and data science. Anyone of yourself, from engineers, account executives, finance, support, run analytics on data from our data warehouse without filing a ticket or waiting on the data team. Nothing I'm gonna do. Users can ask simple questions like, how many leads we got from the campaign, and these are our writes and runs based on SQL. But, importantly, EZero is also a data scientist.
If a user needs statistical analysis, EZero will spin up a sandbox that you just heard about, will run Python, will execute that Python, will generate reports, and spin it all down. D0 is the most used internal tool ever sell. So we have about 800 employees ever sell. answers 30,000 questions a month. And it's safe and standard. D0 doesn't run on Godmode. Every query is sent to the user's permissions.
Athena picks accounts. It plans the outreach. It tracks signals. It runs the weekly motion for every account executive at Roussat. When we launched Athena, pipeline coverage doubled. Every single account executive at Roussel uses a PNode every single day. Under the hood, Athena needed the same kind of semantic layer D0 needed, plus durable workflows and secure connections, and again, a UI this theme keeps coming up.
Because in this case, go-to-market agents are more effective with pixels and buttons than just Slack channel. Third, lead agent. This is our autonomous SDR. We took our best SCR, our best performing SCR, and we trained the lead agent on their playbook. And now the agent runs that playbook autonomously on its own 24/7. You may have seen some headlines.
drafted by draft zero earlier this year, that we took 10 SDRs per cell down to one.
We redeployed those nine routes into bigger roles. We raised our quotas, and we see the 32XRY.
And it costs $5,000 a year to run.
It performs at the 90th percentile of our reps and one engineer manages it part-time. On the hood, lead agent runs on the resilient stack.
Everything CID has told you. AI SDK, workflow SDK, and chat SDK for Slack. And importantly, lead it in as open source. You can find me at hr.gov. No clone for Endulgerium.
Today, we're making a real Maldiva. So I'm going to send an analysis for cell for Enterprise apps are dangerous. It's the same Rostel developer experience you know and love for everyone at your company. Identity and access built in. and the option to run in your phone AWS tenant.
Even if you can control bills with AI, still have to limit access to the absent agents that they filled. And that's why we built Vercel Passport. Shout out to our brilliant Bell here in Berlin and the Vercel County. The Zeldon Passport puts every internal admin agent behind your ID key by default. Internal tools stay internal, employees can only see the apps that they need, and nothing is publicly exposed by default and from day zero.
And you can deploy all this to your own AWS tenant.
You heard that right. You can even monitor cell functions in your own AWS account. Whatever you build with agents stays inside your security boundary. So we believe and we know from experience that the teams who ship fast, who ship securely, who ship at scale are the ones who are going to win. We can build enterprise apps and agents so that can be used. If I can be helpful, please reach out. Otherwise, multi-bacune wrap up.
And again, because this is a bit of a frozen over moment, you can earn out of this on your own AWS account and you can start today. Ask your coding agents to install the Vercel plugin. Then you can build anything. A website, an app, an agent, and ship it at global scale. And your cell agent will keep an eye on production. This is a genuine infrastructure. We can't wait to see what you ship next. Thank you so much. I thought that was quite impressive, you think?
workflow is easy ⇒ workflow sdk
WordpressDK ⇒ wordpress SDK?