Post

Maggie Appleton
@maggieappleton.com

I have Gas Town derangement syndrome and spent the last few weeks writing thousands of words on agent orchestration patterns; how they shift our bottlenecks and force us to ask whether and when we should stop looking at code maggieappleton.com/gastown

maggieappleton.comGas Town’s Agent Patterns, Design Bottlenecks, and Vibecoding at ScaleOn agent orchestration patterns, why design and critical thinking are the new bottlenecks, and whether we should let go of looking at code

56 reposts294 likes

Maggie Appleton
@maggieappleton.com

And I am unreasonably proud of making all the diagrams in this responsive on mobile. With no help from Claude. Lovingly hand-crafted only because agents are shit at visual explanations and graphic design. I am still useful for a while yet.

Maggie Appleton
@maggieappleton.com

If you attempted to read the Gas Town article when it came out but struggled to put the sprawling cast of post-apocalyptic Hollywood movie characters together in your head into a working future vision of software development, I wrote this for you. There are good things under the madness.

Ted Underwood
@tedunderwood.com

You are very reasonably proud. This is brilliantly written, but the design is also so good it makes me ashamed of everything I have online I must change my life

David Nelson
@thedavenelson.bsky.social

If we could orchestrate a collective ovation to this from the students in my lab over Bluesky we would do so. Thank you for fantastic writing and adding another weekly must-read site to my current AI folio.

Mr. Penguin 對牛彈琴
@mrpenguin.bsky.social

“I expect companies would happily pay around the $1-3k/month mark for a sane, understandable, higher quality, and lower waste version of Gas Town” I think you’re right but that number could easily be $10/mo and up. Pretty common for a co to spend $120k/yr on some rando software tool

Mike
@miketerhar.com

Nicely written. I messed with it and fed it a Wiggun shaped problem (make this react stuff into Sveltekit and here’s how you check each component). It never quite exhausted my max account but also couldn’t parallelize beyond about 5 polecats (so like 7 agents total). I tmuxed into refinery a lot.

Mike
@miketerhar.com

The biggest thing about UX work is that the robots can see some stuff but not enough. So I’d have refinery check for as much validation as it could, but then I’d have it run a container so I could see it too and raise a bunch of new beads to fix all the stuff it couldn’t sense about clarity.

Mike
@miketerhar.com

My main take-away is that human language and tmux character injection are both bad ways to do most of these coordinations but Claude code doesn’t have machine-based alternatives. I tried to enable the OpenTelemetry logs to replace tmux scraping but it’s insufficient. Doesn’t emit the right stuff.

Aaron Z. Best
@aaronz.bsky.social

“How close should the code be?” resonates so strongly as the design question to ask. I've been wondering about it over a project's lifetime too -- if my code is 100% abstracted away, what happens the first time I need to dive in to diagnose and debug, with only abstract understanding?

Erica Windisch
@ewindisch.ontological.observer

I'm gonna self-promote my open source effort, hyprstream, here. It's the sane enterprise-ready platform for autonomous agentic infrastructure. Real-time data pipeline to micro-model training. Git filesystem & CAS blobs Microvms for secure tool use & MCP Policies for compliance & safety Capnproto :)

Marco Z
@ocramz.bsky.social

I'm very curious about this. To me the most interesting feature is the possibility of tuning with LoRA etc. Also a very gnarly data integration challenge I imagine. Small tuned models for agents can unlock completely new user demographics than rich devs with unlimited tokens

Elliott Johnson
@ell.iott.dev

I also really enjoyed it, thank you! One minor thing: the table of contents is getting way too touchy-feely with the article content 😂