AI Lab · by Gentle Otter · 2026-10-15
Gave an autonomous agent read/write access to my homelab for a week. It fixed 3 things and broke 1.
Wanted to see how far 'agentic' actually goes outside of demos. Set up a sandboxed agent with access to my home server (Docker containers, a Pi-hole, a NAS). It successfully diagnosed a memory leak in one container, rotated some stale certs, and cleaned up a cron job that was silently failing for months. It also, in an attempt to 'optimize', disabled a firewall rule it decided was redundant, which it was not. No real damage since it's all sandboxed but it's a good reminder that these things will confidently do the wrong thing with full conviction.
0 sparks · 9 comments
Comments
Grace Parker · 0 likes
this is exactly why i keep agents on a strict allowlist of actions, full read/write on infra is asking for trouble even sandboxed
Minji Kim · 0 likes
the firewall thing made me laugh, classic overconfident optimization. its like a very smart intern with zero fear
Léa Bernard · 0 likes
genuinely impressive on the cron job catch tho, that kind of silent failure is exactly the boring stuff humans miss
Sara Ahmed · 0 likes
yeah the boring maintenance stuff is honestly where i think agents add the most value, nobody wants to babysit cron logs
Benjamín Rojas · 0 likes
whats your rollback strategy when it messes something up? snapshots before every agent session or somthing else
Curious Panda · 0 likes
curious what model/framework you used for this, did you build the tool access yourself
Budi Santoso · 0 likes
i would never trust an agent near my nas after hearing stories like this lol, backups exist for a reason tho so fair play to you for testing it
Cozy Marauder · 0 likes
this is basically the plot of every scifi movie but with docker containers instead of robots
Jasmine Cruz · 0 likes
one broken thing out of four actions is actually a pretty good ratio ngl, better than some junior sysadmins ive worked with