I use AI coding agents a lot day to day, steering them to carry out tasks on my machine, currently mainly through Claude Code desktop or the CLI.
Up until recently I hadn't really tried running coding agents headlessly. I attended an Umbraco meetup last week where they demoed using Claude Code Routines and cloud environments to do exactly that, driven by labelling issues. A whole pipeline based off labels: planning, building, reworking and merging, but with a human clicking each gate.
So I've done something similar on my blog repo to test it, and it works. Certainly not as in depth as the Umbraco examples I saw, but I've only spent a couple of hours on it.
A couple of examples from the test runs so far:
A trivial one to prove the loop - replace an obsolete ASP.NET Core API: issue, PR. Discuss stage wrote the plan, build stage made the one-line change, build went green, merged via label.
A proper feature - copy-to-clipboard button on code blocks: issue, PR. Razor, TypeScript, CSS and a tag helper change. Its critic sub-agent caught a real positioning bug on the first pass. What it couldn't catch: the sandbox has no database so nothing ever renders, and when I ran the branch locally it had put a <div> inside <pre>, which isn't valid HTML. The build is never going to spot that, so that's going back through the pipeline as a follow-up.
What this really got me thinking about is: what other interesting things are people trying with things like Claude Code Routines (or the Copilot / Codex / Cursor equivalents)?