Safeguards

Safeguards are the highest-level controls in the platform. While guardrails evaluate the content of individual actions, safeguards control the agent's autonomy and your reactive interventions at a system level.

All of them live under Policies in the sidebar, in the Governance group, each on a page of its own:

• Pauses and Escalation policies — reactive controls that fire while the agent acts
• Execution controls — preventive controls the agent reads before it acts: execution authority (trust levels) and decision thresholds (confidence and data freshness)

Create policy on the Policies page opens each of them: Pause agent actions, Route to a human, and Set execution controls. Each page's breadcrumb leads back to Policies.

Pauses (Emergency Stops)

Navigate to Policies > Pauses (/workspace/policies/pause). The page lists every active pause and the ones lifted recently.

A pause is an emergency stop: it immediately halts the agent's new matching actions for a given scope. Use one when something is going wrong and you need to stop the agent without waiting for individual actions to finish. You can create a pause yourself, and a circuit-breaker pause is created automatically when a monitored metric (such as error rate, rollback rate, anomaly rate, budget burn rate, or SLA-breach rate) crosses its threshold.

To pause everything right away from any page, use Pause agent in the top bar. It opens a short pause dialog with the scope set to all actions.

Creating a Pause

1. Click Create pause (on the Pauses page, or Pause agent actions under Create policy)
2. Choose what the pause does: Pause a skill stops matching actions outright; Route to human review lets them through only after someone approves them, and stops anything with no reviewer to reach
3. Set the Scope — Everything (global), an Action type, a Specific action, or an Entity such as a campaign — and its value
4. Set Expires after — 30 minutes, 1 hour, 4 hours, 24 hours, or "Until manually deactivated"; the default is a 1-hour expiry
5. Add a Reason (recorded in the audit trail)
6. Optionally, click Test against last hour to see how many recent actions the pause would have affected
7. Click Deploy pause

An active pause is enforced immediately: any new action within its scope is blocked, or held for approval if it routes to human review. A pause set to auto-expire stops blocking once its expiry passes (enforced at evaluation time — the underlying record stays on file for the audit trail); a pause with no expiry blocks until someone ends it with End pause, giving a reason.

A pause stops everything in its scope. To block only the actions that match a condition, write an error-severity rule on Blocking rules (Guardrails); to cap spend, use Spend limits (Budget). The builder's Need something else? panel links to both, and to Escalation policies.

Templates

Templates are reusable pause setups that pre-fill the builder with what the pause does, its scope, and its expiry. Once you've chosen what the pause does, name the setup and click Save as template; apply a saved one from the Templates panel beside the builder. Templates are for pauses only — they do not apply to escalation policies.

Active Pauses Indicator

While any pause is active, the top bar shows N pauses active on every page. Select it to open Policies > Pauses. See Activity for the rest of the top-bar status.

Escalation

Navigate to Policies > Escalation policies.

Escalation policies watch a running execution (an automation or chat session) and pull a human in when it starts to behave unsafely, rather than gating a single action up front. Each policy has a type, a scope (* for global or a specific target), a config, and an enabled flag. The available policy types are:

When a policy fires, the execution is escalated with a reason and reason code for review.

The sensitivity gate is a configurable policy type, but it is not yet wired to a data source: its accumulated sensitivity score stays at zero because nothing currently feeds data-classification labels into a running execution, so the gate never fires in practice today. It is reserved for a later phase. The other four policy types are active.

Execution Authority (Trust Levels)

Navigate to Policies > Execution controls. The page has two sections: Execution authority, covered here, and Decision thresholds, covered below.

Trust levels control how much the agent can do without asking. There are three levels:

You can set defaults and scope them, so different tools or contexts can run at different levels.

Decision Thresholds

These settings are the Decision thresholds section of Policies > Execution controls; the old /workspace/policies/thresholds address opens it. The section groups three related settings: Confidence weights, Data Freshness, and Circuit Breaker Thresholds.

Thresholds define the minimum confidence score an action must have to proceed without human review. An action below the threshold is held for review in the Queue.

Data Freshness

Data freshness feeds the confidence score: when a data source the action depends on is stale (older than its configured maxAge), confidence is penalized. Configure which sources matter and how stale they may be here.

Confidence Weights

The confidence score breaks down into dimensions, and you can tune how the score is computed. The dimensions are:

Circuit Breaker Thresholds

This section configures the metric thresholds that trip an automatic circuit-breaker pause — for example, the error rate or budget burn rate at which the agent is halted.