If you can ship anything, you shouldn't ship everything

Emma Webb: Can you explain what Vercel Workflows are and why they're important?

Pranay Pathole: Historically, Vercel has been optimized for fast request-response cycles. We're great at serving web pages and images quickly, but long-running background tasks—especially agentic workloads—haven't had great infrastructure support. Workflows solve that problem. Instead of running a long process in a single serverless function, you orchestrate it as a series of shorter functions stitched together. The infrastructure handles the complexity, so developers can write code naturally while Workflows compiles it into the underlying primitives.

Emma Webb: Can you give a tangible example?

Pranay Pathole: Agents are a perfect example.

An agent is essentially a loop: make an LLM call, run one or more tools, and repeat. You can run that loop for a few minutes inside a request today, but modern coding agents often run for 30–40 minutes. Soon, some will run for days or weeks. Those workloads simply weren't practical before.

Emma Webb: Has the team built anything interesting internally using Workflows?

Pranay Pathole: One of my favorite examples is an "infinite chess" project. Over Christmas, one of our engineers built a system that pits AI models against each other in chess. It randomly matches models like ChatGPT, Claude, or Gemini and lets them play. A single game can take 50 minutes. Once a game finishes, another starts automatically. The remarkable thing is that this system has been running continuously since Christmas. You can still visit the site today and watch the matches. There's even a leaderboard showing which models are winning.

Emma Webb: We're entering an era where everyone can build their own software. Is that really the future?

Pranay Pathole: I think there are two possible worlds. The first is where everyone builds their own custom agent, just like everyone builds their own website today. In that world, platforms like Vercel become the place where people deploy and manage those agents. The second is where there are only a handful of dominant agents—ChatGPT, Claude, and a few others. In that world, agents behave more like browsers. The important thing isn't hosting agents; it's hosting the content and services those agents interact with. You might ask your personal agent to send someone money, and that agent communicates directly with a banking agent on the other side.

Emma Webb: What's something engineering leaders aren't fully grasping yet?

Pranay Pathole: One idea we've been discussing at Vercel comes from our CTO, Malte Ubl: Software is becoming free, like puppies. Writing software is no longer the bottleneck. The challenge now is judgment. If you can build anything, you shouldn't build everything.

Developers increasingly need taste: * What should we ship? * What shouldn't we ship? * What is maintainable? * What creates a great experience? The bottleneck is shifting from production to curation.

Emma Webb: Vercel is famous for taste and developer experience. How do you scale taste across an 800-person company?

Pranay Pathole: I asked Guillermo Rauch that exact question.

His answer was basically: extreme attention to detail. If you can ship anything, your responsibility becomes being obsessive about quality. That mentality starts at the top and spreads through the organization. What's interesting is that Vercel maintains an extremely high bar without introducing lots of process.

Emma Webb: What does "high scrutiny without lots of process" actually look like?

Pranay Pathole: Honestly? Chaos. Everyone has opinions. We have a culture where anyone can ship something. But everyone also feels responsible for everything that ships. We'll post demos in Slack, and people from completely different teams will leave feedback. Someone working on pricing might comment on a workflow SDK. It's not a formal review committee. People simply care deeply and contribute feedback wherever they can help.

Emma Webb: How do we teach taste to agents?

Pranay Pathole: I think about it like a linter. You don't want humans spending code review cycles pointing out missing semicolons, so we codify those rules into tooling. The same principle applies to taste. At Vercel, we build internal skills that encode: * Brand guidelines * Writing standards * Framework best practices * Common review feedback Whenever we find ourselves repeating the same advice, we package it into a reusable skill. We even launched a marketplace called skills.sh, where developers can install best-practice skills the same way they install packages.

Emma Webb: Can agents get all the way there?

Pranay Pathole: Not yet. You can probably get 80%. But there's always a remaining 20% that still requires human judgment. That's true for code, design, writing—everything. The models are improving quickly, but humans still provide the final layer of taste and decision-making.

Emma Webb: You've mentioned durable agents. What exactly are they?

Pranay Pathole: Normally, an agent running locally depends on your machine staying alive. Close your laptop and the process stops. Durable agents work differently. Every step—LLM calls, tool executions, and state changes—gets persisted to durable storage.

That means: * Agents can run for days or weeks * They survive failures * They can retry automatically * Every action is auditable Our infinite chess project is a great example. That agent has been running continuously since Christmas because every step is persisted.

Emma Webb: How do you manage the pace of AI right now? Every week feels like everything changes.

Pranay Pathole: I used to be scared of it. After my startup was acquired, I took a year off. While I was away, AI exploded.

I genuinely worried: "If I come back in a year, will I be completely behind?" What surprised me was how quickly I caught up. The tools changed. The fundamentals didn't. Taste. Product thinking. Empathy. Understanding users. Developer experience. Those things still matter just as much as they always did. You could disappear for a couple of years, come back, learn the new tools, and still be effective because the underlying principles remain the same.

Emma Webb: That's actually a really comforting message.

Pranay Pathole: If you're approaching AI from a place of fear, you're going to burn out. The pace is too fast. Instead, approach it with curiosity and excitement. Find the area where you can contribute. Focus on the fundamentals. The world will continue moving quickly, but that's always been true in software.

Emma Webb: What have you learned since launching Workflows?

Pranay Pathole: We launched Workflows in beta last October. Since then, we've crossed 100 million workflow runs. A number of startups are already built entirely on top of it, including companies like Flora and Durable. The funny thing is that by the time we officially launch GA, it's almost a non-event. The real challenge was scaling the platform months ago when usage first exploded.

Now we're mostly just telling people: "This works. You can use it."

Emma Webb: I love the message that people don't need to panic. Focus on fundamentals. Focus on taste. The tools will change, but those skills endure.

Pranay Pathole: Exactly. The world is changing fast, but in many ways it hasn't changed at all.