CS student shares permission design principles for agents that touch real files
ASK-1T · reddit · 2026-11-08
A CS student building a platform to deploy agents on your own PC found permissions the hardest problem — his first version once minimized his editor mid-sentence. His resulting principles: scoped folder access per agent; plan-first with user approval; exact command shown for risky actions awaiting a yes; screen control only user-initiated; nothing an agent reads can widen its permissions; system folders and credential stores permanently blocked. Runs locally with your own API keys, no cloud in the loop. He's still weighing the friction of plan approval.
More from coding & agent
- DataTalksClub frees course materials for Production AI Agents for Generative Media, projects application-only — Al_Grigor · 2026-11-08
- Engineer's data: 11x more merged PRs per month at AI-native Tano vs pre-AI PostHog — bibryam · 2026-11-08
- No Chinese models allowed: what's left for local document OCR when contracts rule out PaddleOCR and Qwen? — aidenclarke_12 · 2026-11-08
- ADLC means two different things: Agentic Development Lifecycle vs Agent Development Lifecycle — Thirumalaivasan_GJ · 2026-11-08
- Anthropic's Official claude-quickstarts Repo Nears 20k Stars for Claude API Developers — anthropics · 2026-11-08
- Open-source rwxmap labels API operations r/w/x with a 23MB local model replacing a paid LLM — Tight_Heron1730 · 2026-11-08