Vercel Workflows is now GA

VVercel
Computing/SoftwareSmall Business/StartupsBusiness NewsInternet Technology

Transcript

00:00:00- Hey, good morning, everyone.
00:00:02Thanks so much for joining us today.
00:00:03I'm so excited for what we're about to share.
00:00:06So for a decade now, code running on Vercel
00:00:10was primarily built for immediate web requests.
00:00:14But the most interesting software our customers create today
00:00:17no longer fits that model.
00:00:20And the number of these new projects is exploding.
00:00:23Just in the last few months,
00:00:25weekly deployments in Vercel have doubled.
00:00:28And a third of those new deployments come from coding agents,
00:00:31a number that's gone up a thousand percent
00:00:34in just six months.
00:00:35But the most interesting part to me
00:00:38isn't how much we're deploying, it's what we're deploying.
00:00:41Those agents are shipping agents.
00:00:44And the agents that our customers build,
00:00:48they don't just respond once and wait.
00:00:50They need to reason, call tools, wait on humans,
00:00:53and then do it all again.
00:00:55And each of those workflows needs a backend that can keep up.
00:00:58So what I've seen in Vercel
00:01:02is that the team shipping the fastest right now,
00:01:05they aren't rebuilding infrastructure
00:01:07and coding queues themselves,
00:01:08or writing retry logic from scratch,
00:01:11or even wiring of observability.
00:01:13The stuff's been done.
00:01:15So they're using our infrastructure primitives,
00:01:17the stuff we've spent the last decade building,
00:01:20so they can focus all of their focus on their product.
00:01:24And that's exactly the problem we built Workflows to solve.
00:01:27We launched it in beta six months ago,
00:01:30and the response was incredible.
00:01:33And now, a hundred million runs later,
00:01:36we're so excited to announce that Workflows is GA, today.
00:01:41Let me show you how it works.
00:01:43The app I'm gonna show you
00:01:44is a birthday card generator that I built.
00:01:47Now, it could be anything really,
00:01:49but this hits all the things we just talked about.
00:01:51It doesn't fit into a single request,
00:01:54it waits on humans, it retries when things fail,
00:01:57and it streams progress in real time to the user.
00:02:00So, here's how it works.
00:02:02Let me send myself a card,
00:02:04and I'll start by entering my email.
00:02:05Next, I'll pick a date,
00:02:09and I'm gonna pick the 23rd of April.
00:02:11Now, I like doing bonfires for my birthday,
00:02:14so I'll do a Ocean Beach bonfire,
00:02:17and I'm just sending it to myself, so for Pernet.
00:02:21And I'll enter my own email here just to make the demo easy.
00:02:24So, it's gonna have myself RSVP to my own birthday.
00:02:27Cool, now I'm gonna hit Generate and Send,
00:02:30and I'll see the birthday card start generating.
00:02:34I'll head over to the Vercel dashboard
00:02:35where I can see this run just came in.
00:02:37Let me click in.
00:02:39And I can see this in real time as it generates prompts,
00:02:42and then generates the message,
00:02:43and generates an image as well.
00:02:45This is the observability we were talking about.
00:02:48Every step is visible right here.
00:02:50You can see exactly what's running,
00:02:52what's completed, and what's waiting.
00:02:54Over on the user side,
00:02:57I can see that the card's been generated,
00:02:58and the RSVP emails have been sent.
00:03:01At this point, the workflow is suspended.
00:03:04There's no compute actually happening.
00:03:07It's waiting for a human to respond,
00:03:09and I could go into my email, respond right now,
00:03:12or even wait for a few hours or days and respond then.
00:03:15This will continue to track the progress.
00:03:18Now, before I show you the email,
00:03:20I actually wanna show you the CLI.
00:03:22Workflow ships with the CLI on day one
00:03:25that lets you inspect everything that's available
00:03:27on the observability dashboard.
00:03:29This means that you and your agents
00:03:32have access to the complete picture
00:03:35when you're trying to debug something that went wrong.
00:03:37Okay, over to my email.
00:03:40I should see the email come right in.
00:03:43There it is.
00:03:44And I can, it's just asking me to RSVP.
00:03:46I'll hit yes, that'll be there.
00:03:49Back to the dashboard, I can see the hooks resumed,
00:03:53and I should kick off into a sleep.
00:03:55Back to the user dashboard, and I'll see the same thing.
00:03:58We're no longer waiting for RSVPs,
00:04:00and now it's just waiting for the 23rd,
00:04:01so it can send out the birthday.
00:04:03Now, I don't want to make you wait for two weeks,
00:04:05so I'll head back to the dashboard
00:04:06and wake up the sleep right here.
00:04:09That should immediately resume the workflow,
00:04:12and we'll see the workflow complete.
00:04:14Let me head back to the user dashboard.
00:04:17Postcard's been sent.
00:04:18And over in my email, I should get a birthday card
00:04:22with the RSVP as well.
00:04:23Cool.
00:04:24Now, I want to show you the code,
00:04:26and remember the thing we said earlier.
00:04:29Teams shipping the fastest today aren't dealing with queues
00:04:32and managing state and infrastructure.
00:04:34That's the old way.
00:04:35In workflow, everything we've seen today
00:04:38is a single function,
00:04:40and it starts with this directive, useWorkflow.
00:04:43This one directive is what lets this function run for days,
00:04:48suspend between steps, and resume exactly where it left off.
00:04:52So when the workflow calls a step,
00:04:53like generate prompts or generate image,
00:04:56the workflow suspends the function,
00:04:58queues the work in the background.
00:05:00It'll even retry it if it fails,
00:05:02and then resume the workflow when it's done.
00:05:05The entire time, I get full observability
00:05:07on the Vercel dashboard of every step along the way.
00:05:11I can even click into one of these,
00:05:13and by default, on Vercel, everything's encrypted,
00:05:16but I'll decrypt the data so you can see the input.
00:05:19Yeah, this is available both on the dashboard and the CLI.
00:05:22Workflows also come with built-in streaming.
00:05:24This is how I pushed real-time updates to the client
00:05:29to track progress live.
00:05:31I didn't have to pull in some sort of database
00:05:33just to pull this off.
00:05:34The client can even disconnect and reconnect just fine.
00:05:37This is a persistent stream.
00:05:39All right, and here's the hook.
00:05:42For every RSVP, I'm creating an ephemeral webhook URL
00:05:46on the fly, I'm then sending it in an email,
00:05:50and then I can simply just wait for it down here.
00:05:52The workflow pauses right here
00:05:56until every person clicks on their link.
00:05:59There is no database involved,
00:06:01and the results even just live in local JavaScript variables.
00:06:05The workflow holds all of its own state.
00:06:08This is the distributed systems machinery
00:06:11that we said was just gone.
00:06:12You don't have to think about this stuff anymore.
00:06:14All right, lastly, let's sleep at the end
00:06:17to sleep for my birthday.
00:06:18It's right here.
00:06:19This sleep could be a couple of seconds,
00:06:21days or even weeks like we just did.
00:06:24And while it's sleeping, the workflow is just suspended
00:06:27with zero compute running.
00:06:29When it wakes up, it continues exactly where it left off,
00:06:33maintaining state along the way.
00:06:34And this is the entire thing.
00:06:37It's one function that you can just read top to bottom
00:06:40and reason about from end to end.
00:06:43Every step is backed by a queue,
00:06:45it's retried automatically,
00:06:47and runs on its own serverless function.
00:06:49There are no timeout limits,
00:06:51so workflows can run as long as they need to.
00:06:54This is what we mean
00:06:56when we say Vercel handles everything underneath.
00:06:59You can just write the logic so we take care of the rest.
00:07:02Now, in this demo, I showed you a simple use case,
00:07:07but our customers use workflows
00:07:09to run agents and production at tremendous scale.
00:07:12Mux uses workflows to ensure video generation pipelines
00:07:17run reliably every time.
00:07:19And Flora is orchestrating parallel creative agents
00:07:23across 50 image generation models, all with workflow.
00:07:27Durable is a small team
00:07:30that hosts 3 million customer websites,
00:07:33and they use workflows to ship new agents
00:07:35to their users every single day.
00:07:38Now, I started this demo
00:07:40by saying that the most interesting software
00:07:42our customers build no longer fits in a single request.
00:07:46And we're seeing more and more apps like these
00:07:50deployed every single day.
00:07:51Agents are shipping agents,
00:07:54and the teams building them aren't thinking about queues,
00:07:58they're focusing on their customer problems.
00:08:01And the shift isn't just happening inside Vercel.
00:08:05Our customer deployments are doubling,
00:08:07and fewer engineers than ever are handrolling infrastructure.
00:08:11So whatever app or agent you want to build,
00:08:16workflows will make it reliable,
00:08:18long-running, durable, and observable.
00:08:22From your laptop to a million users
00:08:25all with the same application code.
00:08:27Workflows is generally available today,
00:08:31so go shift something that works.
00:08:37[BLANK_AUDIO]

Key Takeaway

Vercel Workflows transitions long-running logic into a single durable function that manages state, retries, and human-in-the-loop interactions without requiring external queues or databases.

Highlights

Vercel Workflows is now generally available after reaching 100 million runs during its six-month beta period.

Weekly deployments on Vercel have doubled recently, with one-third of new deployments originating from coding agents.

Coding agent deployments on the platform increased by 1,000% over the last six months.

The useWorkflow directive enables a single function to run for days or weeks by automatically suspending and resuming state between steps.

Each workflow step runs on its own serverless function with automatic retries and no timeout limits.

Ephemeral webhook URLs can be created on the fly within a function to pause execution until a human interaction occurs.

Companies like Mux and Flora use Workflows to manage video generation pipelines and orchestrate parallel creative agents across 50 different models.

Timeline

The Rise of Agentic Software and Coding Agents

  • Modern software no longer fits the traditional model of immediate web requests.
  • Coding agents now account for one-third of all new deployments on Vercel.
  • Agentic workflows require backends capable of reasoning, calling tools, and waiting on human input.
  • High-velocity teams utilize infrastructure primitives instead of building custom retry logic or observability.

The landscape of web development is shifting toward complex, multi-step processes rather than simple request-response cycles. Deployment volume is surging, driven largely by autonomous coding agents that have seen a 1,000% growth rate in six months. These agents build other agents that must maintain state over long periods, necessitating a shift away from hand-rolled infrastructure like manual queues and complex observability wiring.

Observability and Human-in-the-Loop Interaction

  • Workflows provide real-time visibility into every step of a long-running process through a centralized dashboard.
  • Execution suspends entirely with zero compute cost while waiting for human responses via email or webhooks.
  • Progress updates stream live to the client without requiring a dedicated database for state management.

A birthday card generator serves as a practical example of a workflow that spans multiple days and involves human interaction. The Vercel dashboard displays live progress as the system generates prompts, messages, and images. When the system sends an RSVP email, the workflow enters a suspended state, consuming no compute resources until the user interacts with the generated link.

CLI Integration and Workflow Resumption

  • The Workflows CLI allows developers and agents to inspect the state and data of any run for debugging.
  • Manual overrides can wake a sleeping workflow to resume execution immediately.
  • Resumed workflows maintain all previous state and local variables upon waking.

The command-line interface provides the same level of observability as the dashboard, making it accessible for both humans and AI agents to debug failing steps. During the demo, a workflow scheduled to sleep for two weeks is manually woken through the dashboard to demonstrate immediate resumption. This flexibility allows for testing long-duration logic without waiting for real-time delays.

Technical Implementation and useWorkflow Directive

  • The useWorkflow directive manages the suspension and queuing of background work automatically.
  • Local JavaScript variables hold the results of distributed steps without an external database.
  • Data is encrypted by default on Vercel but can be decrypted for inspection in the dashboard or CLI.
  • Persistent streams allow clients to disconnect and reconnect to a running workflow without losing progress.

Workflows simplify distributed systems by allowing complex logic to be written as a single, readable function. When a step like image generation is called, the function suspends and the work is queued; once finished, the function resumes exactly where it left off. This model eliminates the need for managing database state for simple tasks, as the workflow engine handles the underlying state persistence and encryption.

Production Use Cases and Scale

  • Mux ensures reliable video generation pipelines by using Workflows to manage sequential tasks.
  • Durable supports 3 million customer websites by shipping new agents daily via Workflows.
  • Flora orchestrates 50 different image models simultaneously using parallel workflow logic.

Beyond simple demos, companies are using Workflows to solve high-scale production challenges. Mux relies on the system for video processing reliability, while Flora uses it for massive parallel orchestration across diverse AI models. This GA release marks a shift where engineers can deploy durable, long-running applications from a laptop to millions of users using the same unified code.

Community Posts

View all posts