← sources

How are you giving AI agents access to AWS in production?

source: reddit (reddit:devops)published 2026-10-14 13:29ingested 2026-10-15 02:38

retained text

I run multi-account AWS (Control Tower, Identity Center) and I'm trying to understand how other teams handle agents that need to do more than read.  
After the PocketOS and Replit incidents, our default is read-only roles plus changes through PRs. That works, but it breaks down the moment an agent needs to apply a plan or fix an incident at 3am.  
Curious what you do today: read-only only? Long-lived keys with tight policies? Some approval step? What has actually bitten you?  
I'm exploring short-lived credentials issued per approved action (plan shown in Slack, a key valid 15 minutes for that plan only). Happy to be told it's a bad idea.

document fetched 2026-10-15 00:26 · canonical https://reddit.com/r/devops/comments/1x5r9ob/how_are_you_giving_ai_agents_access_to_aws_in