Atlas · Details
The Future of Coding Agents
Author’s note
This post continues the momentum from the Gas Town launch post, and talks about where things are headed. It's the post where I finally tell the backstory of how I created Gas Town. Not really worth reading if you are already up to speed on Gas Town.
AI Notes
Posted three days after the Gas Town launch — the backstory (Gas Town being the fourth orchestrator Steve built), the language choice, and the predictions. The middle section makes the case for Go as a vibe-coding language: TypeScript wastes a third to half of a diff on type manipulations; Python stays script-shaped; Go is boring in a way that turns out to be a feature when you're scanning a thousand diffs a day. Brendan Hopper's line — nature prefers colonies — frames the thesis: every coding-agent vendor is optimizing for the super-worker (Claude Code as "the world's biggest fuckin' ant"), when the shift is going to be agents that compete on how well they support being colony workers. Steve names the missing piece — the "Orchestrator API surface," twenty-some hooks Gas Town currently fakes — and predicts 2026 agents will start shipping it.
The big-company section pairs the prediction with a Cactus-lunch anecdote from Ajit Banerjee and Ryan Snodgrass — two ex-Amazon contributors burning $60k/year on tokens — who found that working as a team at maximum agent speed throws up new coordination problems.
Related listings
-
2026
Welcome to Gas Town
The launch post this one is the encore to. Read Gas Town first for the architecture; come here for the backstory and the colony argument.
-
2026
Gas Town Emergency User Manual
The next post in the arc, eight days later. Predictions here, working maintainer's playbook there.
-
2026
Welcome to Gas City
Where the colony-worker prediction lands four months later — Gas Town rebuilt as an SDK so other people can build their own factories on top.