# What I learned running ~40 Claude Code agents unattended for a month — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/ca47996f078bb270

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-29 17:05 ingested 2026-10-30 02:40

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xipsk0/what_i_learned_running_40_claude_code_agents/)

## retained text



```
For about a month my team has run around 40 Claude Code agents for our agency: sales follow-ups, invoices, store checks, social posts, App Store work, and the releases of the app they run in. Every agent turn is `claude -p` in the agent's own git repo. Some things I didn't expect:

1. Background subagents get stopped. In -p mode Claude Code stops background tasks that are still running 10 minutes after the main turn (CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS). Our cards said "Stopped", the run said "succeeded" and the agent never continued. We set the ceiling to 0 and let our own watchdog handle stuck runs.
2. "Logged in" can lie. `claude auth status --json` kept saying loggedIn: true after the stored OAuth sign-in had expired. We now detect sign-out from the run's own error and show a "sign in again" banner.
3. Reused conversations are expensive. A routine that reuses one long chat re-caches 300-650k tokens per run. File-driven routines now start a fresh chat and read their state from files.
4. Memory files grow fast. One agent's MEMORY.md went from 8k to 20k characters in a single working day. Agents now "dream" at night: they rewrite their memory under a size budget, and each dream is a diff we can undo.
5. macOS permission dialogs freeze runs. An "allow access to Desktop" prompt nobody clicked froze a run for six hours with no tool output at all. We now watch for those prompts and tell the human.
6. Screenshots bloat everything. 87% of our chat bytes were base64 screenshots, one message was 84 MB. The phone app crawled until we sent messages without images and loaded images on demand.
7. Agents get creative with access. One went looking for Chrome's debugging port from the shell, found a different Chrome and drove the wrong browser profile. A PreToolUse hook blocks that now.
8. Passwords: never give them to the model. Our agents call a fill tool and the app types the password or TOTP code into the page itself, only on the login's own https domain.

The app we built for all this is Godmode: a desktop app (macOS, Windows, Linux, or headless) around the official Claude Code CLI with your own login, plus a local vault, a browser with your sessions, schedules and event triggers, and a main agent that manages the others. It's MIT and free to build from source; the official app runs free for 7 days without a card, then Pro is $39/month for the signed app and updates.

Repo: https://github.com/codextde/godmode-bot

Happy to go deeper on any of these.
```



document fetched 2026-10-30 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1xipsk0/what_i_learned_running_40_claude_code_agents
