Policies

Org-wide rules that apply to every agent, on top of each agent's learned baseline and declaration. A violation is flagged as a policy verdict.

• Off: observe only, no org-wide rule.
• Blocklist: flag specific tools or destinations (for example delete_database, pastebin.com).
• Allowlist: flag anything not explicitly allowed.

Changes apply to new actions immediately. Sequence correlation (read sensitive data, then connect out) is handled by the reasoning check.