Steve Yegge identifies 8 stages of AI-assisted coding adoption.
I’m somewhere between Stage 5 and 6 - I’ve largely abandoned the IDE in my own projects in favor of managing amp or claude code CLI tools (Stage 5). I like living on the command line anyway, that’s why I’ve always been a vim user.
However, I haven’t really graduated to running multiple concurrent agents (Stage 6). I don’t really mind waiting for a single agent to sequentially knock out tasks for personal work, but it’s easy to see the usefulness of orchestrating a team of AI coders for e.g. enterprise software projects with huge monorepos.
At first I was just running claude or amp as an interactive session and guiding it step by step through building my projects, but lately I like using Geoff Huntley’s Ralph Wiggum coding loop, particularly this great implementation for claude/amp by Ryan Carson.
And that’s all it is really, a bash loop. As Addy Osmani points out, what this design gives you is memory persistence, so you don’t have to swamp the agent with the full context of past and future tasks in each iteration; external files - git, progress.txt, prd.json - maintain the state of dev progress and each loop iteration can pull a subset of context and requirements for carrying out a task.
Again, I’m using a single agent here (Yegge Stage 5), and it sets passes: true to each sequential task in the PRD file prd.json as it completes it. This takes time; generally once I have my PRD ready I run the loop (ralph.sh) and walk away from my laptop for a bit. I suppose you could add dependency logic to the PRD’s tasks such that multiple claude/amp instances could be picking off tasks concurrently,
But at any rate this is where Yegge’s Gas Town comes in. This is a tool built on Steve’s earlier project beads, and it abstracts agent work into a well-defined (and colorful) taxonomy. I haven’t gone deep into this yet, expect more to come here.