Building for the Agent Era: Notes from Vercel London

The Lookout is on the 50th floor of 8 Bishopsgate. When you arrive on a March Tuesday evening, the City of London spreads out in all directions in the way that cities only do from height: everything small, everything connected, the river a silver thread through the middle of it. It's the kind of venue that front-loads a mood.

Vercel had rented it for the evening to host "Rewriting the Rules: Building for the Agent Era." About 150 people, by my rough count. Mostly developers and engineering leads, some product people, a few who looked like they'd wandered in from financial services and were trying to work out what to make of the canapés. The usual post-work event demographic: laptop bags, slightly too-casual-for-the-venue trainers, lanyards that should really have been left in the bag.

I went because the premise interested me and because the speaker list was genuinely good. I'm writing about it because the central argument of the evening turned out to be more clarifying than I expected.

The Central Argument

Tom Occhino, Vercel's CPO, opened with the keynote framing. Occhino was a key figure behind React at Meta, which gives him a specific kind of authority to talk about how the web's component model evolved. He was inside it when it was being built. The argument he made, roughly, was this: the web has always reorganised itself around whoever is doing the navigating, and the navigating is now being done by agents.

The way he put it (paraphrasing from notes) was something like: UIs are becoming leaf nodes. The agent is the new trunk. You're no longer primarily building for a human who clicks through a series of pages; you're building for a software process that needs to read, reason about, and act on your content programmatically, potentially without rendering a UI at all.

This isn't a completely new idea. The zero-click search conversation has been circling this territory for a while, and the agentic commerce infrastructure discussion that's been building since Google's UCP announcement at NRF is essentially the same idea from the commerce side. What Occhino's framing added was the developer-layer perspective: what does it actually mean for how you architect a web product if the primary consumer of your API is an autonomous agent rather than a human holding a phone?

The honest answer, which came up repeatedly across the evening, is that we're mostly not there yet. The architecture for agent-native web products is still being designed.

Alex Holt and the Practicality Problem

Alex Holt from ElevenLabs gave what was, for me, the sharpest talk of the evening. Before ElevenLabs, Holt was at Palantir where he worked on NHS COVID data infrastructure during the pandemic. That gives him a particular vantage point on what happens when AI goes into production in genuinely high-stakes environments rather than demo conditions.

The point he kept returning to was the gap between "an agent can do this" and "an agent can do this reliably, in production, at scale, without the failure mode being catastrophic." That gap is, in most real systems, still substantial. The NHS work was an object lesson in what happens when data quality is poor, when edge cases outnumber common cases, and when the consequences of errors are measured in something other than conversion rate.

What this translates to for commerce and web products specifically: the path from "we built an agent that can do X" to "we shipped an agent that does X for our customers" is longer and harder than the demo phase suggests. The infrastructure for observability, fallback handling, and graceful degradation is the part most companies haven't properly invested in yet. That's the operational reliability work nobody in a demo bothers to show you.

There was a specific thing he said about trust that stuck with me: "The agent era requires the user to trust the agent's judgement at the moment of action, and that trust is earned slowly and destroyed quickly." Which is, if you think about it, pretty much the design brief for every checkout flow ever written, translated into agent terms.

Stuart Brown on M&S's Reality

The fireside chat brought in Stuart Brown, Head of Software Engineering at M&S, and Jim Drury, Head of Frontend Technologies at Virgin Media O2. Two enterprise engineers talking honestly about what deploying agents into real systems actually looks like.

Brown was the one I found most useful to listen to, partly because M&S operates at the intersection of physical and digital retail in a way that makes the agent conversation genuinely complex. You have store systems, you have a significant online business, you have a loyalty programme with serious data depth, you have brand integrity considerations that a fast-moving DTC startup doesn't have to worry about in the same way. The M&S personalisation and loyalty work I've been tracking has been substantial, but the engineering questions around deploying agents that interact with customers in high-trust contexts (a customer asking about a garment, a customer trying to resolve an order issue) are different from deploying agents that optimise internal workflows.

The honest version of what Brown was describing was: we're learning to walk before we run. Agents in internal developer tooling, agents in content pipeline automation, agents in the test-and-monitoring layer. These are the deployments that are building the organisational muscle memory for how to work with agentic systems. The consumer-facing deployments are coming, but the groundwork matters.

Jim Drury's perspective from Virgin Media O2 was complementary: the frontend architecture questions are real, and the industry is slightly behind on working out how to build UIs that degrade gracefully when the primary interaction mode is an agent that may or may not render your carefully constructed component tree.

What I Took Away

Walking out into the March evening with a glass of something that had been pressed into my hand during the networking, the thing I kept turning over was the disconnect between the ambition of the framing (agents as the new trunk, UIs as leaf nodes, the web fundamentally reorganising around autonomous processes) and the stage we're actually at, which is mostly "we've got some agents doing useful things in controlled conditions and we're working out how to make that robust."

That's not a criticism of the evening. The ambition is necessary. You have to build toward something in order to make infrastructure decisions that will age well. The companies that are thinking about agent-native architecture now are the ones that won't be caught flat-footed when the agent interaction model becomes more prevalent. The Alibaba Accio story is worth keeping in mind here: 230,000 businesses on a no-code agentic commerce platform since its launch. That's one data point for how fast adoption can move when the tooling gets accessible enough.

But Holt's point about the gap between demo and production is the honest anchor. The agent era is real and it's arriving. It is not, as of March 2026, fully here.

The view from the 50th floor was better than anything inside. Worth the trip up, though.