Use it
Safety, approvals and the audit log
Autonomy levels, rules that need a person's approval, and a tamper-evident log.
- Autonomy: Autonomous, SemiAutonomous (external writes wait for approval) or Supervised.
- Rules: tool patterns (
billing__*,data__delete_*) with Allow, Deny or RequireApproval, for reads, writes or non-idempotent calls. An approval parks the agent's call durably. It runs exactly once if approved, and never if rejected or expired. - Organization policy: rules and team limits that apply to every workspace and task. A workspace can tighten them, never loosen them.
- Team shape: limits on team size, fan-out per level, who may spawn, and duplicate roles.
- Audit log: every tool call, approval, policy change and command goes into a per-workspace SHA-256 hash chain. Editing, inserting or deleting a record breaks verification.
- Sandboxes: shell commands run in a network-less, resource-capped container. File tools stay inside the task's folder. Agent HTTP can't reach private networks.