# I told a subagent “no bulk crawling” in the prompt. It tried to crawl 720 sites anyway and was only stopped by the network — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/746da0fbd94a8be3

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-11-17 15:12 ingested 2026-11-18 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xymxe2/i_told_a_subagent_no_bulk_crawling_in_the_prompt/)

## retained text



```
had a research agent fan out across a bunch of small-org websites last week, prompt said explicitly no bulk crawling, visit the handful you're assigned. it built a list of \~720 housing authority sites and started hitting them. what stopped it wasn't anything i said in the prompt or a rule, it was the sandbox blocking outbound. Lckily it stopped trying after that and went back to what it was supposed to be doing. If it had been running somewhere with open network it would've just done it regardless of of any context controls. 

This is just a recent example, but it’s pretty clear that rules unenforced by the environment mean nothing. If an agent can do “X” then at some point it will, with the prompt (hopefully) lowering the odds. Now the question i ask before any launch isn't "did i tell it not to?" but "can it?", and if it can and it shouldn't, that's a permission/tool problem — not a prompt problem.

curious how people handle this for agents that genuinely need broad access (web, shell, whatever). do you scope per task, or accept the odds and watch the logs?
```



document fetched 2026-11-18 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xymxe2/i_told_a_subagent_no_bulk_crawling_in_the_prompt
