Vercel thinks agents are a genuinely new form of software, not a fresh coat of paint on the web apps it made its name shipping. That’s the core claim from Andrew Qu, the company’s Chief of Software, in an interview with Latent Space. Qu works alongside Vercel’s CTO on internal engineering and emerging tech, and he led the development of eve, Vercel’s new framework for building agents. His argument matters because Vercel sits at the infrastructure layer where a lot of the web actually gets built.
What stands out here is the framing. Qu doesn’t call agents “apps with an LLM bolted on.” He calls them a different category with different rules.
Why agents count as new software
Qu’s reasoning is practical, not philosophical. “They are not as predictable as web applications,” he told Latent Space. “The infrastructure can look similar, but the interaction, interface and outputs are much more dynamic.”
That unpredictability forces new building blocks. According to Latent Space, Vercel hit a wall of “paper cuts” while building v0, its vibe-coding product: switching models and providers, adding fallbacks, making runs resumable. None of the existing tooling solved these cleanly. So Vercel turned its fixes into reusable primitives and bundled them into eve.
The primitives Qu names are worth writing down, because they’re becoming the shared vocabulary of agent engineering:
- Context handling and compaction
- Tools and filesystem access
- Resumability for long-running work
- Subagents
- Skills as portable, on-demand knowledge
The part most teams underestimate: skills
Qu makes a sharp point about why skills have become so important. Models ship with stale knowledge. His own example: agents still recommend Vercel Postgres, a product the company deprecated years ago. A skill can correct that on the fly, steering the model toward the current approach.
His phrase for this is “forward-correct the model.” Until a company can audit and update every old blog post, doc, and tutorial the model trained on, skills act as a patch layer. The practical takeaway is direct: publish skills for the latest version of your product, then go audit your existing content and flag what’s outdated.
This is a real to-do item for any business whose docs feed into AI answers. If the model is confidently recommending your dead product, that’s a support and revenue problem hiding in plain sight.
When to let the agent run, and when to babysit
Qu resists the two loudest camps in the agent debate. He doesn’t think the future is fully autonomous loops, and he doesn’t think it’s all human-in-the-loop either. “It is about choosing a feedback cycle that fits the task,” he said.
His rule of thumb is useful. If the task is well defined and you know what the output should look like, let the loop run to completion. For surgical engineering work, check back in and keep steering. The best candidates for agents, per Qu, are repetitive tasks that still need reasoning: contract redlining first passes, marketing retrospectives, writing queries against internal data.
The bigger signal: the web is being rebuilt for bots
The most forward-looking claim comes at the end. Vercel has published data showing bot traffic rising while human traffic stays flat or declines, even as impressions climb. Qu’s conclusion is blunt: “The future of the web is therefore to be as accessible to bots and agents as possible.”
That reframes a decade of web strategy. Sites were built for human eyes and search rankings. Now the fastest-growing visitor is an agent trying to understand and use your product. Vercel already detects when an agent makes a request, which hints at where its platform is heading.
What to do with this
A few practical moves fall out of Qu’s comments:
- Treat agents as a distinct build target, not a feature. Plan for context, tools, and resumability from the start.
- Publish skills for your current products and audit old content that models still parrot.
- Match autonomy to task risk. Loops for defined work, humans for surgical work.
- Make your site legible to agents, not just to people and search crawlers.
Vercel is betting its next platform on all of this, the same way it once bet on making web deployment easy. Whether eve becomes the default the way its web tooling did is the open question. The full interview is worth reading at Latent Space for Qu’s take on how far this shift goes.