Comment by jmspring
1 day ago
I think Steve's idea of an agent coordinator and the general model could make sense. There is a lot of discussion (and even work from Anthropic, OpenAI, etc) around multiagent workflows.
Is Gas Town the implementation? I'm not sure.
What is interesting is seeing how this paradigm can help improve one's workflow. There is still a lot of guidance and structuring of prompts / claude.md / whichever files that need to be carefully written.
If there is a push for the equivalent of helm charts and crds for gas town, then I will be concerned.
I ran into this building a similar workflow with LangGraph. The prompt engineering is definitely a pain, but the real bottleneck with the coordinator model turns out to be the compounding context costs. You end up passing the full state history back and forth, so you are paying for the same tokens repeatedly. Between that and the latency from serial round-trips, it becomes very hard to justify in production.