Real-time Monitoring

Real-time Monitoring evaluates AI activity while a session is running. It combines inline action review with the Watcher agent, which reviews evolving session behavior, maintains a safety state, and can apply configured interventions when risk crosses organization thresholds.

​

Inline review and the Watcher agent

The two controls operate at different points. Inline review is suited to a single action that can be evaluated immediately. The Watcher agent is suited to patterns such as escalating intent, repeated denied activity, suspicious tool sequences, scope changes, or accumulated risk across a session.

​

Enable Real-time Monitoring

1. Open Settings.
2. Select Observability.
3. Find Real-time monitoring.
4. Select an Inline action review mode.
5. Select a Session watcher mode.
6. Choose the appropriate Failure behavior.

Users with policy-management permission can change these settings. Configuration changes are written to the Audit Log.

​

Watcher modes

Suggest context cannot require approval or stop a session. Those actions require Enforce mode.

​

What the Watcher evaluates

A Watcher assessment can consider:

• Newly accepted session events.
• Recent prompts, responses, tool activity, and policy decisions available from the source.
• The current session summary and safety state.
• Earlier Watcher assessments and event judgments.
• Active security signals.
• Risk category, severity, confidence, and evidence continuity.

The Watcher records the events covered by each assessment so reviewers can trace its conclusion back to the relevant session activity.

​

Available interventions

Active stopping must also be explicitly enabled and is disabled by default. Organizations can adjust Watcher thresholds through the real-time governance configuration API. Added guidance expires after 30 minutes. Requiring approval for the next risky action expires after 2 hours unless it is resolved earlier.

​

What stopping means

The Watcher can stop a session only when its activity passes through a supported Forge enforcement point. A stop intervention:

• Blocks the governed session at its next supported blocking event.
• Applies to the identified session rather than every session belonging to the user.
• Prevents future governed activity while the intervention remains active.

It does not terminate an arbitrary operating-system process, undo an action that already completed, or stop activity that does not pass through a supported enforcement point. Because the Watcher runs asynchronously, activity may occur before an assessment finishes. Use inline action review when a decision must be made before an individual action proceeds.

​

Failure behavior

Fail open allows a supported inline action to continue when the real-time reviewer cannot return a decision in time. Fail closed prevents the action from continuing when the reviewer cannot establish a decision. Watcher assessments are asynchronous. If an assessment fails, Forge does not invent an assessment or create a new intervention. Existing active interventions remain enforceable until they expire or are resolved.

​

Review Watcher activity

Open a session to review:

• Current safety state.
• Risk level and confidence.
• Watcher rationale.
• Covered evidence.
• Recommended action.
• Active intervention.
• Inline review decisions.

Qualifying assessments can also create security signals and related findings for investigation. Intervention activation and resolution are recorded as governance and audit evidence. Configuration changes are recorded in the organization Audit Log.

​

Related pages