Vercel Workflows is now GA

Hey, good morning everyone. Thanks so much for  joining us today. I'm so excited for what we're   about to share. So for a decade now, code running  on Vercel was primarily built for immediate web   requests. But the most interesting software  our customers create today no longer fits that   model. And the number of these new projects  is exploding just in the last few months.  Weekly deployments in Vercel have doubled, and a  third of those new deployments come from coding   agents, a number that's gone up 1,000% in  just six months. But the most interesting   part to me isn't how much we're deploying, it's  what we're deploying. Those agents are shipping   agents and the agents that our customers  build. They don't just respond once and   wait. They need to reason. Called tools.  We run humans and then do it all again.   And each of those workflows needs a backend  that can keep up. So what I've seen in Vercel   is that the teams shipping the fastest right  now, they aren't rebuilding infrastructure and   coding queues themselves, or writing retry logic  from scratch, or even wiring of observability.  The stuff's been done. So they're using our  infrastructure primitives, the stuff we've   spent the last decade building so they can  focus all of their focus on their product.  And that's exactly the problem.  We built workflows to solve.  We launched it in beta six months ago,  and the response was incredible. And now,   100 million runs later, we're so excited  to announce that workflows is GA today.  Let me show you how it works.  The app I'm going to show you   is a birthday card generator that I built. Now it could be anything really, but this hits   all the things we just talked about. It doesn't  fit into a single request. The waits in humans, it   retries when things fail and it streams progress  in real time to the user. So here's how it works.  Let me send myself a card, and I'll start by  entering my email. Next, I'll pick a date.  And I'm going to pick the 23rd of April.  Now I like doing bonfires my birthday,   so I'll do a potion beach bonfire and  I'm sending it to myself. So for any.  And I'll enter my own email here just to  make the demo easy. So it's going to have   myself RSVP to my own birthday. Cool. Now I'm  going to hit Generate and Send and I'll see   the birthday card start generating. I'll head  over to the Rowsell dashboard where I can see   this run just came in. Let me click in. And I can see this in real time as a   generates prompts and then generates the  message and generates an image as well.  This is the observability you were talking  about. Every step is visible right here.  You can see exactly what's running,  what's completed and what's waiting.  Over. On the user side, I can see that the  card's been generated and the RSVP emails   have been sent. At this point, the workflow  is suspended. There's no compute actually   happening. It's waiting for a human to respond.  And I could go into my email, respond right now,   or even wait for a few hours or days and respond.  Then this will continue to track the progress.  Now, before I show you the email, I actually  want to show you the CLI workflow ships with. The   CLI on day one that lets you inspect everything  that's available on the observability dashboard.  This means that you and your agents have access to  the complete picture. When you're trying to debug   something that went wrong. Okay, over to my email. I should see the email come right in. There it is.   And I can. It's just asking me to  RSVP. I'll hit yes, that'll be there.  Back to the dashboard. I can see the hooks  resumed and I should kick off into a sleep.  Back to the user dashboard and I'll see the  same thing. We're no longer waiting for RSVP,   and now it's just waiting for the  23rd so I can send out the birthday.  Now I want to make you wait for  two weeks, so I'll head back to the   dashboard and wake up the sleep right here. That should immediately resume the workflow,   and we'll see the workflow complete when we  head back to the user dashboard. Postcards   been sent and over in my email I should  get a birthday card with the RSVP as well.  Cool. Now I want   to show you the code and remember the thing we  sent earlier. Teams shipping the fastest today   aren't dealing with queues and managing state and  infrastructure. That's the old way in workflow.   Everything we've seen today is a single function  and it starts with this directive. Use workflow.  This one directive is what lets this  function run for days, suspend between   steps and resume exactly where it left off. So when the workflow calls a step like generate   prompts or generate image. The workflow  suspends the function queues to work in   the background element, retry if it fails,  and then resume the workflow when it's done.   The entire time. I get full observability on the  Vercel dashboard of every step along the way.  I can even click into one of these and by default  on Russell, everything's encrypted, but I'll   decrypt the data so you can see the input. Yeah, this is available both   on the dashboard on the CLI. Workflows also come with built in streaming.  This is how I pushed real time updates  to the client to track progress live.  I didn't have to pull in some sort of  database just to pull this off. The   client can even disconnect and reconnect  just fine. This is a persistent stream.  All right. And here's the hook.  For every RSVP, I'm creating an ephemeral webhook  URL on the fly. I'm then sending in an email,   and then I can simply just wait for it down here. The workflow pauses right here until every   person clicks on their link. There was no  database involved, and the results even   just live in local JavaScript variables.  The workflow holds all of its own state.  This is the distributed systems machinery  that we said was just gone. You don't   have to think about this stuff anymore. All right, lastly, let's sleep at the end   to sleep for my birthday. It's right here. The sleep could be a couple of seconds,   days or even weeks like we just did. And while  it's sleeping, the workflow is just suspended   with zero compute running. When it wakes  up, it continues exactly where it left off.  Maintaining state along the way. And this is the entire thing. It's   one function that you can just read top to bottom  and reason about from end to end. Every step is   backed by a queue. It's a retry automatically  and runs on its own serverless function. There   are no timeout limits, so workflows can run  as long as they need to. This is what we mean   when we said Vercel handles everything underneath. You can just write the logic so we take care of   the rest. Now in this demo I showed you a simple  in use case, but our customers use workflows to   run agents and production at tremendous scale.  Mocks uses workflows to ensure video generation   pipelines run reliably every time. And Flora's orchestrating parallel   creative agents across 50 image  generation models all with workflow.  Durable is a small team that hosts 3 million  customer websites, and they use workflows to   ship new agents to their users every single day. Now I started this demo by saying that the   most interesting software our customers  build no longer fits in a single request,   and we're seeing more and more apps  like these deployed every single day.  Agents are shipping agents, and the teams  building them aren't thinking about cuz   they're focusing on their customer problems. And the shift isn't just happening inside   Vercel. Our customer deployments  are doubling and fewer engineers   than ever are hand rolling infrastructure.  So whatever app or agent you want to build,   workflows will make it reliable, long  running, durable, and observable.  From your laptop to million users, all  with the same application code. Workflows   is generally available today, so  go achieve something that works.