permissions

To control AI agent command execution, you can configure explicit allowlists and blocklists using the permissions skill. This enforces security by ensuring agents only run approved commands, preventing unauthorized actions and accidental damage.

Yes, you can require manual approval before an AI agent runs dangerous commands by enabling an approval workflow. The permissions skill manages this by intercepting command execution and holding potentially dangerous operations until explicit approval is granted.

Available security modes for managing tool access policies include deny, allowlist, blocklist, and full execution. These modes enforce different levels of command execution security to ensure deterministic task execution for AI agents.

Tool policies prevent an AI agent from deleting critical files by defining specific tool access rules and grouping tools into profiles. Configuring a blocklist or explicit allowlist restricts access to destructive commands, safeguarding your system.

The best way to group tools for deterministic AI task execution is by using tool profile expansion features. This allows you to define granular rule management and assign specific tool groupings to agents, ensuring strict security controls.

Use a blocklist to explicitly forbid known dangerous commands while allowing all others, or an allowlist to restrict agent execution to only pre-approved safe commands. Choose allowlists for strict security and blocklists for flexible but controlled environments.