
On Friday I was fortunate to be on a call with Thorsten Ball, one of the people behind Amp, while he walked through his own development workflow in it.
Somewhere in the middle of it, it became clear to me that Amp has been moving in pretty much the direction I’d worked out I wanted when I built Kingston Falls. And they’ve done a better job of it than I have.
Amp and Me
Amp was one of the first Gen AI dev tools I used. It was also one of the first to walk away from the IDE. In February Amp announced that the coding agent is dead and that its VS Code and Cursor extensions would “self-destruct on March 5 at 8pm Pacific Time”. They did, and the CLI was what was left.
I kept using Amp on and off in the terminal after that, mixed in with Claude Code, Codex, Antigravity and others, whichever suited the job. I built Kingston Falls for Claude Code, and later extended it to Codex.
What I’d Missed
What I hadn’t noticed was Amp pivoting back to an IDE of sorts, in its own UI.
It isn’t an editor. It’s somewhere to run and watch a lot of agents at once. Each thread runs in its own orb, a sandboxed machine with your code in it, and from the web or the Mac app you get the thread, its diffs, its files, a terminal and even a desktop on that machine. A portal shows you whatever the agent has running there, so you can try the change. Artifacts hold the reports. Puck sits over all of it as a coordinator, and Spaces, which shipped the week before the call, give you one agent whose job is to start and steer the others. When Amp took the thread sidebar out of its terminal UI in August, it said why: “the best place to keep track of all those threads is in the Amp native and web apps”.
That’s the path Kingston Falls took me down too. I started with a terminal and Slack, and while Claude Code was still building the factory I had it build me a pane to watch it with. A web dashboard came later. I couldn’t see what the agents were doing, and I needed to. Autonomous work is only as trustworthy as your ability to watch it, and in the end the watching wanted a proper screen.
The Same List
I gave Puck my description of Kingston Falls and asked it to map how I’d do the same in Amp.
Most of the list came back with an Amp name next to it:
- The hub and runner VMs I keep alive with leases, heartbeats and checkpoints are orbs, which Amp runs. A paused orb keeps its workspace, so there’s nothing to check in to a cocoon branch, and nothing is wiped when I reopen a task.
- The Slack thread for each task is an Amp thread, which I can pick up on the web, the Mac, my phone or in Slack.
- The controller is Puck, or a Space.
- Steering a busy agent without stopping it is how Amp handles a message now: it’s delivered at the next opportunity instead of being queued until the turn ends.
- A Vercel preview for every task branch is a portal.
- Published research reports are artifacts.
- The independent review after every task is a separate review thread. I set mine up so the review always runs on another provider’s model, and a commit after the verdict voids it, as it does in Kingston Falls.
- The fences in each project’s profile are AGENTS.md, skills and plugins.
- The Claude and ChatGPT subscriptions the fleet runs on work in Amp too, and since September Amp has been free to use with your own compute and subscriptions.
It wasn’t a perfect match. What Amp doesn’t have are the things that make Kingston Falls safe against a hostile agent rather than a clumsy one: an egress allow-list, a commit signer outside the sandbox, pushes limited to the task’s own branches, and approval cards that only I can approve. There’s no per-thread budget either. The nearest Amp gets is a Ship setting per project, which I can set to push a branch and raise a pull request, and leave the merge to me.
Seeing so many of the ideas I’d only arrived at by building them turn up in someone else’s product was the best kind of validation. It also made the next step easy.
Moving In
So, for the time being at least, I’m putting Kingston Falls down and immersing myself in Amp, to see how the workflow fits me.
In the same conversation with Puck I set up Spaces for CALM and Roadmapper, each with its own rules for how work ships.
The Question Nobody Has Answered
There was one question Thorsten didn’t have an answer for when I asked him: how do you cope with the constant context switching?
I asked because of a day earlier this month. I flew to New York and ran the factory from Slack on my phone all day, most of the flight included: about nine deploys, a research report and a new project taken to a paid GitHub App by the evening. It felt like being the 10x developer, and it left me drained. Every card was a decision. I was the bottleneck by design, so whenever I wasn’t looking, something was waiting for me. When acting on an idea costs almost nothing, there’s always one more ready. And the work doesn’t stop when I do.
Amp doesn’t change any of that. If anything it makes it easier to have more going on at once: more threads, more orbs, more Spaces. Amp’s pitch for Spaces is “sitting in the captain’s chair”, and that does feel like the right seat. But a captain still makes every call.
I ended my first Kingston Falls post saying the agents do the work now, and the watching is harder to give up. Moving to Amp means I no longer have to build and run the factory. The watching I still do myself, and I don’t have an answer for that yet either.