Crimble Crumble

← Blog

Cloudflare Connect 2026: notes from a conference that was really about AI agents

San Francisco skyline on a clear day, with the Transamerica Pyramid and Salesforce Tower above rows of houses
Nils Huenerfuerst / Unsplash

Last month I was in San Francisco for Cloudflare Connect. On paper it’s a Cloudflare conference. In practice it was a few days of “AI agents are the future of the internet”, and nearly every talk, hallway chat and product pitch came back to that.

The headline that set the tone: according to Cloudflare, traffic from AI agents (software acting on someone’s behalf, rather than a person clicking around) has now overtaken human traffic on the web, and they reckon it’ll keep climbing steeply. Most of the event was about what to do about that.

More humane than I expected

I went in braced for a few days of hype. The talks were upbeat, but there was a real awareness of the downsides too. Plenty of speakers talked openly about what could go wrong and what the industry should be doing to keep it on the rails. That was a nice surprise.

The tooling is further along than I thought

The engineering side is also more mature than I’d assumed. The setups people showed for building, deploying and running AI-driven systems weren’t science experiments. They were proper pipelines with checks, audit trails and guardrails built in.

What stuck

The bottleneck is now people pressing “approve”

The productivity gains everyone talks about only show up once agents can do real work end to end: merge code, change infrastructure, complete tasks we’re currently nervous about handing over. The catch is you need pipelines that check their own work at each stage and keep a clear record of every step an agent took to make a change. Otherwise you’ve got a very fast assistant waiting on a human to approve each step.

Local AI tools should be able to run the whole loop

On a developer’s own machine, the goal is a setup that can plan a feature, build it, review its own work (ideally with more than one AI model checking), fix what the review finds, and then raise it for a human to look at. You shouldn’t need to hit Return every 30 seconds to keep it going.

Lock those local agents down hard

Agents running on your machine shouldn’t have direct access to your important command-line tools, passwords or API keys. The approach that kept coming up was to route their access through a controlled gateway using MCP (Model Context Protocol, a standard way for AI tools to connect to other systems), so you decide exactly what they can reach. I wrote more about moving MCP from pilot to production earlier this year.

Attacks no longer let up

Traditional security assumed attacks would eventually end: someone hammers your site, you ride it out, things calm down. Speakers argued that no longer holds: automated attackers don’t get tired, so you plan for a steady stream instead. The bots are also getting smarter quickly, which is pushing a lot of new detection work.

Telling humans and bots apart is getting harder

When an agent sits between a person and your website, it gets much harder to work out who your actual users are and what kind of traffic you’re seeing. People are working on it, but I think this will be a pain for a while yet.

Don’t rely entirely on someone else’s AI

If a critical part of your business depends on a closed AI service, you’re at the mercy of that company’s pricing, policies and roadmap. Several speakers argued for keeping more control over the models you rely on, and I’m with them.

Cloudflare Workers are cooler than I gave them credit for

I’d always thought of Workers as “a bit of code you can run at the edge” (on Cloudflare’s servers close to the user, rather than in one central data centre). It’s a much bigger ecosystem than that. Pair it with Durable Objects (small units of storage and logic that hold state between requests) and you can build some genuinely clever things.

A few smaller things

  • Blocking dodgy packages before they land. Cloudflare showed how its WARP client can block risky or unknown software packages before they’re installed on a machine. That matters more now that AI coding tools happily pull in packages on your behalf.
  • Agent payments with a spending limit. Mastercard showed an agent payment system I really liked. Money moves from your card or bank account into escrow held by a smart contract (a bit of code that enforces the rules). The agent can only spend what’s been set aside and never touches the funding source directly.
  • AI video has crossed a line. From what was on show, generated video is now very hard to tell apart from the real thing.
  • “Proof of human” is coming. Proving you’re a real person is likely to start with a check in the physical world that then carries over online. World ID and Cloudflare are working together on this: you verify yourself once in front of one of World ID’s camera devices, and later checks happen through your phone’s camera.

Where I’ve landed

I came home more optimistic than I expected, and with a long list of things to dig into. Things are moving fast, but the people building the plumbing seem to be thinking it through, and much of the tooling already exists.

If you’re wondering what all this means for your website or internal systems, have a chat with us about getting AI-ready.