Coding with agents and subagents has some new interesting articles on it...
A lot of them point at markdown-driven Claude Code setups. In this one the author referred to the Ralph Wiggum coding agent
You set up PROMPT.md with this
0a. read everything in specs/
0b. review the files in src/
1. pick the SINGLE highest priority item from @IMPLEMENTATION_PLAN.md and implement it using up to 50 subagents
2. ensure the tests and checks are passing
3. update the @IMPLEMENTATION_PLAN.md with your progress and commit all changes with git add -A && git commit -m "..."
if there is a discrepancy in the IMPLEMENTATION_PLAN.md and the spec, always update the IMPLEMENTATION_PLAN.md to match the spec.
Then you run:
while :; do
cat PROMPT.md | npx --yes @sourcegraph/amp
done
While burning money to run the AI agents.
https://github.com/repomirrorhq/repomirror/tree/main
Original article.
https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md
The key takeaways I get from it, that I agree on:
Be careful... you can sink a lot of money on agenic AI anything pretty quickly...
(oh, and yeah, our team of three is averaging about $12k on opus per month)
As soon as it runs out of context, you are battling a goldfish brain. Better to summarize progress and direction and start with a fresh context
We use a family of techniques I call "frequent intentional compaction" - deliberately structuring how you feed context to the AI throughout the development process.
You can limit context overrun if you take code searching and sign posting of your markdown and give that task to a subagent
The most common/straightforward use case for subagents is to let you use a fresh context window to do finding/searching/summarizing that enables the parent agent to get straight to work without clouding its context window with Glob / Grep / Read / etc calls.
Here's other people's comments on the subject: