# I run my Claude Code agents like a starship crew: engineers in worktrees, a Number One that reviews and deploys, one seat at a time — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/724de12b4ff81595

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-19 12:05 ingested 2026-10-20 02:39

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1x9v2gj/i_run_my_claude_code_agents_like_a_starship_crew/)

## retained text



```
For two months a crew of Claude Code sessions has built my personal journal app (Star Trek styled, single user) while I give orders from my phone. There's a demo you can click around in, on a made-up journal: [https://public.mihai-safta.dev/lcars-demo/index.html](https://public.mihai-safta.dev/lcars-demo/index.html)



The method is the part I think is worth sharing:



\- A worktree per engineer. Each engineer is a Claude Code session (Opus 5.5) in its own git worktree and branch, cut from the live branch, running in herdr so it survives me closing the page. I reach the same session from the app's console, the Claude app or a terminal.

\- A Number One that reviews. A separate session (Fable 5.1) turns my one-line asks into orders, reads every diff and report, and is the only one that deploys. The engineers report to it, not to me.

\- One seat. Only one engineer runs builds and tests at a time; the rest wait in a queue. Parallel agents fought over the same files, and resolving the conflicts burned more tokens than the waiting costs.

\- A deploy gate. Nothing goes live on an agent's word. A fixed script rebases, runs the test suite (including a 60 fps frame budget in headless Chromium), merges and restarts, and rolls back if health fails. 28 of 230 deploys failed, almost all caught by the tests.

\- Estimates. Every engineer gives one before its first edit. They run about 3x too long.

\- Memory in the app, not in files. Anything the crew must remember is an entry in the journal, so I can read and edit it.



[CLAUDE.md](http://CLAUDE.md) does most of the work: build and test rules, where production is, what never to touch. Happy to share any part of it.



Full write-up with videos (including a real navigator agent planning a hike from one sentence): [https://mihai-safta.dev/posts/lcars\_journal/?utm\_source=reddit&utm\_medium=social&utm\_campaign=lcars\_journal&utm\_content=r\_claudecode](https://mihai-safta.dev/posts/lcars_journal/?utm_source=reddit&utm_medium=social&utm_campaign=lcars_journal&utm_content=r_claudecode)



Not open source yet; the demo is the way to see it.
```



document fetched 2026-10-20 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x9v2gj/i_run_my_claude_code_agents_like_a_starship_crew
