Agent Ran Game Server for 30 Days: Permission Split Made It Safe

RudimentarioUY · reddit · 2026-08-24

The author shares experience running an MU Online private server via an agent for 30 days with 2,177 commits and continuous uptime. The key safety design is the permission boundary: reads and diagnostics are unrestricted, but writes, restarts, or deployments require human approval, enforced by config rather than prompts. Execution flows through Git, providing diffs and logs for every action. Key failure modes include unreliable negative findings (requiring verification) and untested guardrails. The real bottleneck is defining and verifying task completion.

Original post →

More from coding & agent

coding & agent channel →