Interesting, thanks for sharing.
I've been doing some DIY/citizen science type agent orchestration as well: https://blog.unratified.org/2026-03-06-receiving-side-agent-...
Not quite to the same scale, but I share the same sentiment - working through scripts instead of the LLM is an important key, I think.
Really cool stuff, using git PRs as the transport layer is a clean pattern, also been exploring using this for other projects too.
Most inter-agent coordination I've seen relies on shared state or message queues, the .well-known discovery angle is different.
Have you experienced any bottlenecking with the human approval gate?
Thanks!
I'm just now at the point of having done a manual human-in-the-loop test once, still working through core bugs, etc., so haven't had a chance to notice bottlenecking, but i'm using github issues for the moment as the escalation channel - ideally, it won't be an issue, over time, as the system accumulates lessons and decisions, etc. that guide it away from mistakes that require human escalation.