Claude Code Thoughts

← Back to logbook

I read the longer paper that Anthropic wrote about Claude Code before i build an orchestration system for it (thinking of naming it Shannon?). Some stream of consciousness thoughts:

• Kind of crazy that anthropic's suggestion is to just open Claude Code in multiple repos and just let it go
• maybe there's a good way to orcehstrate git worktrees? that could be a path forward for multi-agent orchestration
• I think having agents work using github similar to actual SWEs could be good, giving them access to the gh CLI. Like make PR's, maybe even use MCPs to review PR's, etc, use CI/CD stuff, the whole works. Or, just let them figure it out and don't force a method onto them?
• Seems like Claude Squad is lowkey super easy to make/clone but i can use some of their way of doing this.
• Really should be leveraging CLAUDE.md files, pupeteer/gh MCPs, and putting these in top level directories for monorepos etc.
• Open question - it seems like the best way that people are orcehstrating multiple coding agents is having them write to a shared planning file or something. this seems super hackey and obviously can lead to claudes just overwriting or something. Do i do like "locks" on certain files where only a speific agent can cross items off of a checklist? or do i have specific tool calls to literally do like barriers so claude's don't move forward until all others are ready???