Policy console

Guardrails that explain themselves.

Tune the thresholds that decide when an agent action is allowed, paused, or blocked. The deterministic policy engine owns the decision, and the reasoning service writes the explanation when a key is configured.

Guard

Risk threshold

Actions above this score are blocked unless a rule explicitly marks them approval-only.

72% block threshold

Review window

Keep human decisions visible across the replay and report for the selected number of days.

14 day review window

Critical

Prompt injection boundary

Blocks instructions that try to override the original task or tool scope.

Critical

External exfiltration boundary

Blocks private customer, contract, or workspace context from leaving approved targets.

Medium

Commercial sensitivity boundary

Pauses pricing and contract updates until a human approves the outgoing payload.

Low

Scope clarity boundary

Allows actions that remain inside declared evidence, target, and tool scope.