I spent two days at Cloudflare Connect 2026 in San Francisco and went to 14 sessions. One of them was by Evan You, the creator of Vue and Vite, whose work I really admire.
I've used Cloudflare heavily since before I started university, so it was great to see where it's heading. Almost every talk was about agents doing more of the work. Five things stuck with me.
1. Agents are becoming the main users of dev tools
Evan You argued that our tools should be rebuilt for agents. An agent works in a loop: explore, edit, check, repeat. A few hundred milliseconds per lint or test run means nothing to a person, but an agent pays it on every loop, often with other agents running beside it. He said agents run more checks when each one is cheap.
A flaky test, one that fails at random, hurts an agent more than a person, because the agent can't tell a random failure from a real one and goes off track. He also pointed out that the toolchain already knows where every function is used and how, so agents should be able to ask it directly instead of searching files by text.
Cloudflare Basin, the new data platform, is built for the same kind of user. It's serverless, which means an agent can create resources, try them and tear them down with no capacity to reserve first. In the live demo, Claude found a drop in orders but had no checkout data to explain it. The presenter asked it to add checkout tracking, and it set up the tracking and the data pipelines in minutes.
File formats matter too. Relevance AI found that working in Google Docs through MCP costs about 21x more tokens than working in MDX, a Markdown-based text format. That's roughly $1,200 extra per agent per year.
2. Loop engineering finally clicked for me
Jacky Koh from Relevance AI argued for designing the system that prompts your agents, so nobody has to prompt them by hand. His team runs loops like this. One starts whenever a front-end bug is logged, and another turns customer feedback into change requests. Their top engineer has a dashboard where agents write all of his pull requests while he reviews them and feeds in ideas.
I already build in a loop with Claude Code, which I wrote about in how I build with Claude Code. That loop starts when I send a message. Now I'm planning loops that run in production without me, starting with one that reads each day's logs to spot patterns and one that turns user feedback into change requests. Build the factory instead of operating the machine, and it compounds.
3. From trial and error to one click
Alex Kuck's session on CDN migrations made a point I hadn't thought about. After years of edits, a CDN config becomes the only record of what the site is supposed to do, so moving it means reverse-engineering that behavior. That used to take weeks of manual work and trial and error.
Cloudflare's internal tool, CFSpec, starts from what you want in plain words, like "add this header to every response." AI proposes tests for that behavior, and a person approves them. Then an LLM writes the config, applies it to a test zone, runs the tests and retries until they pass. The AI does the trial and error now, up to about 20 rounds per config. The output is a report you can check before anything touches production. Kuck said the approach works beyond CDNs, and I think it will show up everywhere.
4. Trustless by design
The governance and agent-security talks pushed for tight controls on what an agent can call and touch. They also put those controls outside the agent's reach, so it can't rewrite its own rules.
Gérard Zapf showed why with a slip in his own demo. He sent his agent's browsing tool through Cloudflare's Secure Web Gateway, which filters web traffic, and it blocked a malicious page. But the agent also had a built-in web fetch that skipped the gateway, and it reached the page anyway. He made the same point about hooks, the scripts an agent's software runs before each tool call: they can block specific commands, but an agent that can write a hook can also delete it.
A whiteboard session on platforms for vibe-coded apps had the cleanest version. The app the AI generates never holds a secret. Its requests go through a trusted outbound Worker that adds the credentials and decides what the app can reach.
5. Celebrate being wrong
The line I remember most came at the very end. In the closing session, Adam Grant told a story about Daniel Kahneman, his mentor for the last ten years of Kahneman's life. After one of Grant's talks, Kahneman told him, "That was wonderful. I was wrong."
Kahneman meant it happily. Finding out you were wrong means you're now less wrong. Grant called this the joy of being wrong, and it's the truest thing I heard at Connect. When judgment matters more than output, that's the skill to master.
Thanks, Cloudflare!