Back to ai safety
Trust

Agent Permission Models

Compare how agents handle file access, command execution, and human approval.

safetypermissionsagentsecurity

Common patterns

Pattern Description Examples
Read-only Agent can inspect files and logs but cannot write or run commands Static analysis tools
Suggested edits Agent proposes diffs; human must apply them GitHub Copilot, many code review agents
Command preview Agent generates shell commands; human approves each Claude Code, Cline
Auto-approve safe ops Low-risk commands run automatically; destructive ones gated Cursor agent mode
Full auto Agent runs commands and commits without approval; highest risk Some CI agents

Recommendations

  • Start with read-only or suggested-edit modes.
  • Use command preview for any agent touching production systems.
  • Log every action for audit review.