Gas Town: From Clown Show to v1.0

The important audit at my company is conducted by the FDA.

I have a feeling when they ask what processes we followed to mitigate any user harm that could be caused by software changes that "I told an AI-mayor in the form of a cartoon fox what to do and he spit out a bunch of vibecode software written by AI-driven virtual cartoon characters" is not among the answers they want to hear.

Let's assume that managing context well is a problem and that this kind of orchestration solves it. But I see another problem with agents:

When designing a system or a component we have ideas that form invariants. Sometimes the invariant is big, like a certain grand architecture, and sometimes it's small, like the selection of a data structure. Eventually, though, you want to add a feature that clashes with that invariant. At that point there are usually three choices:

* Don't add the feature. The invariant is a useful simplifying principle and it's more important than the feature.

* Add the feature inelegantly or inefficiently on top of the invariant. Hey, not every feature has to be elegant or efficient.

* Go back and change the invariant. You've just learnt something new that you hadn't considered that put things in a new light, and there's a better approach.

Often, only one of these is right. Usually, one of these is very, very wrong, and with bad consequences.

But picking among them isn't a matter of context. It's a matter of judgment and the models - not the harnesses - get this judgment wrong far too often (they go with what they know - the "average" of their training - or they just don't get it). So often, in fact, that mistakes quickly accumulate and compound, and after a few bad decisions like this the codebase is unsalvageable. Today's models are just not good enough (yet) to create a complete sustainable product on their own. You just can't trust them to make wise decisions. Study after study and experiement after experiment show this.

Now, perhaps we make better judgment calls because we have context that the agent doesn't. But we can't really dump everything we know, from facts to lessons, and that pertains to every abstraction layer of the software, into documents (and even if we could, today's models couldn't handle them). So even if it is a matter of context, it is not something that can be solved with better context management. Having an audit trail is nice, but not if it's a trail of one bad decision after another.

It is pretty magical to go from brainstorming an idea in the evening, having ChatGPT Pro spit out a long list of beads to implement it, leaving it running over night in a totally empty repo and waking up to a mostly-implemented project.

So I slapped together my own Beads implementation (https://codeberg.org/mutablecc/dingles) over a day or two. Probably has bugs, and I'm sure race conditions if you tried to use with Gas Town, and likely does not scale. But it has the minimum functionality needed to create and track issues and sync them (locally and remotely, either via normal merge, or a dedicated ticket branch). No SQL, no extra features, just JSONL and Git. Threw a whole large software project at it, and the AI took to it like a duck to water, used it to make epics for the whole project, methodically worked through them all, dependencies first, across multiple context sessions. The paradigm of making tools the AI wants to use is clearly a winner.

At this point it should be clear that Gas Town has done something we can evaluate the value of.

I wound up building my own with Claude, I made it SQLite first, syncs to GitHub, can pull down from GitHub, and I added "Gates" to stopgap Claude or whatever agent from marking things complete if they've not been: compiled, unit tests run, or simple human testing / confirmation. The Gates concept improved my experience with Claude, all too often it says it finished something, when in fact it did not. Every task must have a gate, and gates must pass before you can close a task. Gates can be reused across tasks, so if "Run unit tests" is one gate, you can reuse it for every task, when it passes, it passes for that one task <-> gate combination.

Anyway, I'm happy for Beads, Gas Town not so much my wheelhouse on the other hand.

I think Yegge's instincts that making a programmable / editable coordination layer (he calls this gas city) is a great idea. Gas town early days was definitely a wild experience in terms of needing to watch carefully lest your system be destroyed, and then I put that energy into OpenClaw - I'll probably spin up Gas City and see what it can do soon though. Very cool.

So now you have agents of type mayor, polecats, witnesses, deacons, dogs etc plus a slew of Unneeded constructs with incomprehensible names.

In one of the blog post for gas town I remember reading something by the author along the lines of « it’s super inefficient, but because you burn so many tokens, you still get what you want at the end! » clearly this is also the design philosophy behind this project, just (get your ai to) throw more random abstractions and more agent types until you feel like it kinda works, don’t bother asking yourself if they actually contribute anything.

This gave me the very clear feeling that most of the complexity of gas town is absolutely not needed and probably detrimental.

Ended up building my own thing that is 10x simpler, just a simple main agent you talk to, that can dispatch subagents, they all communicate, wake each other up and keep track of work through a simple CLI. No « refinery » or « wasteland » or « molecule » or « convoys » or « deacons » or …

Seems like I'm back to obscurity.

:)

I tried tracking down where those numbers came from and the sources were a bit sketchy. Can anybody who has used Gas Town confirm those numbers, or report their personal numbers?

This is a desirable end state for highly social but perhaps slightly sociopathic extroverts who want to spend all day talking even though they aren't talking to a person.

For anyone else, it's hard to imagine considering that a desirable way to spend eight hours a day.