Back to full roadmap
topicadvanced
Computer Use Safety
Autonomous agents misclick — sandbox + approval + audit + revert capability mandatory.
2 hours1 prereqs
Computer use = the riskiest production pattern. Mandatory layers:
- Sandbox VM — agent accesses an isolated VM/container, not the host OS
- No real credentials — test data + test accounts; zero agent access to production data
- Action approval gates — user prompt before destructive action ("Should I send this email?")
- Site allowlist — agent operates only on permitted domains
- Audit log + replay — screen recording + click log for every action
- Kill switch — user can say "STOP" at any time
- Rate limiting — max X actions/sec, prevent brute-force behavior