Have been playing around with coding agent orchestrators recently, notably Paperclip and Gas Town, just to see what they’re capable of. They’re interesting modes of working in their own respect.

Paperclip, recommend to me by a work colleague, is a bit more polished and has quite a strange approach to interacting with the agents. Basically, you create pretend organisations, with a business goal, org chart, projects, etc. You as the human take on the role of the “board”, and you give instructions to a “CEO” agent who then “hires” a “CTO” agent to do the coding work. This can branch out to other agents taking on other “roles” like marketing or documentation writing, all which do work dispatched either by those further up the org chart or by you, the human. Interacting it feels like buying into the illusion: I found myself writing tickets along the lines of “the board will request…” or “the board has approved…”

Gas Town, by Steve Yegge, is a bit more of a scrappy upstart. The installation process was less polished and when launched, is little more than a Tmux session in-front of a Claude Code instance that has been heavily customised. Based on how Yegge describes it, it feels more automated in how it organises the agents and work, which I took to mean that you can take a more hands off approach, to the extent you feel comfortable. This has not been my experience so far, but to be fair, I’ve only had two interactions with the Mayor so far. What it does seem to have is community. Yegge talks about the group of developers that have taken his idea and have ran ahead with it, spinning it out into things like Gas City which is a project used to build other orchestrators.

As to what these orchestrators could actually product, well, I guess it depends on your tolerance of vibe-coded artefacts. So far, I’ve only used these to produce toys and prototypes; nothing that could be considered “serious”. I don’t know if it’s possible to actually make anything really commercial with these. I’m afraid to try myself.