Ship London was yesterday. I am writing this on a slightly over-caffeinated Thursday morning with the conference lanyard still in my bag and a fresh agent/ folder on disk.

I do not want another keynote recap. I want the practitioner question: what does "Next.js for agents" actually change in a TypeScript dev house tomorrow?

Vercel launched Eve as a filesystem-first framework for durable agents. The punchline is almost rude in its simplicity: an agent is a directory.

No bespoke orchestration novel. No "platform team" ceremony before you can say hello. The same mental model Next.js taught the web: conventions over glue, local first, production path included.

That is the bit worth stealing even if you never touch the rest of the launch deck.

Frontend engineers already think in folders. Routes are files. Components are files. Config is files. We review directories in PRs. We diff markdown. We argue about names.

Eve leans into that prejudice. Instructions live in instructions.md where a human can read them. Tools are TypeScript modules with Zod schemas, not JSON blobs someone edited by hand at midnight. Skills load like the skill files we already scattered across .agents/. Channels are just more modules.

If you have suffered through "the agent prompt is in this one Notion doc, trust me", you can feel the relief in your shoulders.

The other half of the pitch matters once you leave the demo: conversations as durable workflows. Checkpointed steps. Crash, resume, redeploy without orphaning a mid-flight task on the old revision.

That is the production path Next.js folks recognise. next dev honesty, then vercel deploy without rewriting the architecture for "real" traffic. Eve wants the same deal for agents: honest on a laptop, still true when the URL is public.

I care because most agent demos die the moment you need approvals, retries, or a Slack thread that lasts longer than a coffee.

Get agent/agent.ts and agent/instructions.md honest. Add one typed tool. Talk to it over the HTTP session APIs. Only then start inventing multi-agent empires.

The anti-pattern from the web years still applies: people who skip the boring directory conventions end up rebuilding Next.js badly. People who skip Eve's directory conventions will rebuild agent platforms badly, with extra YAML.

Inside VMO2 we have been circling a knowledge specialist that needs sandboxing, a durable ephemeral filesystem, Slack, and a proper web UI. A framework that treats the agent as a reviewable directory is exactly the substrate for that kind of product, not a chat toy.

I will write that story properly next. Working title in my notes: Katie Knows. Product name, not a pronoun joke.

For this morning, the takeaway is narrower.

Agents got their app/ directory. Everything else was overdue paperwork.