I gave myself a team of AI engineers that open their own PRs | Rafal Muszynski

I gave myself a team of AI engineers that open their own PRs<br>7 Jun, 2026<br>• 8 min read

I have more tickets than time.

I have a day job and a product of my own on the side, where I test things. The product has a backlog, and most of it is small — a fix here, a tweak there. Not junior work, exactly. The kind of thing I’d hand to another developer if I had one — someone who’d do it the way I would. I don’t. So the tickets sit.

Last weekend I tried something. I already lean on Claude Code for most of my actual coding, but that still means I’m the one sitting there. What if I could assign a ticket the way I’d assign it to a person, and have it come back as a pull request I review? Not on my laptop, not while I babysit it — in the background, while I do something else.

I built it. It works. The agent has a name: Dude. You assign a ticket to Dude in Linear, and a while later a pull request shows up on GitHub with the change, a description, and a link back to the ticket.

This is the story of building it, which is mostly the story of everything that was harder than “the AI writes code.”

The whole loop. I plan and file the ticket; everything after "assign" runs on Anthropic's infra, not my laptop.

What it actually is

Dude runs on Anthropic’s Managed Agents, and the mental model took me a second to get straight: I don’t run the agent. Anthropic runs it — the loop that decides what to do next, and the container where the code actually executes: the cloning, the editing, the pushing. I hand off a ticket and wait for the result. What I host is a thin glue service that catches a webhook when I assign a ticket, kicks off a run, and gets out of the way.

That glue is around a hundred lines of Python on a tiny Fly machine — the kind of small webhook bot I’ve wired up before. The cheapest box they rent, a shared-cpu-1x with 256MB of RAM, $1.94 a month — it doesn’t need to be more than that, since the heavy lifting happens on Anthropic’s side, not here. Assigning a ticket to Dude fires a webhook; the glue catches it, works out which repo or repos the ticket touches, starts an agent session pointed at them, and hands it the ticket. When the agent finishes, a second webhook pings me. No stream to hold open, nothing running on my machine.

The “which repo” part is just a label — when I set one. I made a Linear template so every ticket I file for Dude comes out the same shape, and the template can carry a label — backend, admin-ui, www, docs — that pins it to a single repo. Set it and Dude stays there; leave it off and Dude works out which repos the change needs. So the whole act of dispatching work is: pick the template, write the ticket, maybe set a label, assign it to Dude. The label narrows where it works; assigning to Dude decides that it works.

A ticket assigned to Dude — a label for the repo, the assignee that triggers the run, and the PR it opened linked right back on the ticket.

That part took an afternoon.

A Dude session, running on Anthropic's side — reading the ticket, checking the repo, working the task. I don't sit here while it happens; I just read it afterward.

The wiring took the rest

Then I spent the next couple of hours on the unglamorous 90%.

The demo of one of these agents is a five-minute thing. It reads a ticket, edits files, opens a PR. Watching that the first time is the part that makes you think it’s done. It is not done. Done is when it reliably does the job with no human in the loop, and the distance between those two is entirely plumbing.

None of it was hard in the way writing code is hard. It was the slow kind — auth, scopes, configuration that had to line up exactly across two systems. Each wall cost more time than it had any right to and taught me nothing I’ll ever reuse. That’s the 90%.

It opened a PR

Eventually it did the whole thing on its own. A branch named after the ticket. A draft pull request with a summary and a link back. I reviewed it; it was a reasonable change.

A pull request Dude opened on its own, with its own What/Why writeup, from a ticket I'd assigned and walked away from.

There’s a specific feeling to reviewing a PR you didn’t write, for a ticket you forgot you’d assigned.

How I use it now

The planning still happens with me in the loop, which is the part I’m least willing to give up. For anything ambitious I open Claude Code, work the approach in plan mode until I’m happy with it, and have it write the ticket straight into Linear. The plan I’d have written anyway becomes the brief. Then I assign it and walk away.

What I didn’t expect to lean on as much: I can fire several at once. Each assigned ticket spins up its own isolated run, so a handful work in parallel while I do something else. That turned out to be the real gain — not that I’ve stopped babysitting one agent, but that five are working at the same time. On a stack of small, independent tickets that’s an easy 10x. It...