Is anyone using decision models as a permission checker?
retained text
Don't run with dangerous full permissions, unless you are prepared to let the agent fuck your shit up. So says conventional wisdom. But also, I am definitely running with full permissions. When I give my agent a goal to work at things overnight, I necessarily need to give it permissions. So the obvious answers were sandbox and backups. But why not use decision models? Elevation requests come with a justification and the command the main agent wants to run, that seems like plenty of information for a decision agent to give a good Y/N response to vet things. IDK how effective this would be, anyone already doing something similar and can share exp?
document fetched 2026-11-15 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xvxziy/is_anyone_using_decision_models_as_a_permission