Risk Engine

The risk engine evaluates every tool call against your policies and assigns risk scores. It runs synchronously — tool calls are evaluated in real-time before the agent proceeds.

How it works

Tool call received
       │
       ▼
┌─────────────────┐
│ Load all enabled │
│ policies         │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Match conditions │──→ No match → ALLOWED (no risk)
│ against tool call│
└────────┬────────┘
         │ One or more match
         ▼
┌─────────────────┐
│ Apply strictest  │
│ action wins      │
└────────┬────────┘
         │
         ├─→ block           → Tool call BLOCKED, risk event created
         ├─→ require_approval → Tool call PAUSED, approval request created
         ├─→ warn            → Tool call ALLOWED, risk event flagged
         └─→ allow           → Tool call ALLOWED

Risk severity levels

Level	Description	Example
low	Minor concern, informational	Read access to non-sensitive files
medium	Moderate concern, worth monitoring	Write operations to staging
high	Significant concern, may need review	Access to sensitive files
critical	Immediate attention required	Destructive commands, production deploys

Risk events

When a policy matches, a risk event is created with:

• Severity — based on the policy action
• Policy ID — which policy triggered
• Tool call details — what the agent tried to do
• Action taken — block, warn, or require_approval
• Timestamp — when it was evaluated

Risk events are visible in:

• The Runs detail page (timeline view)
• The Dashboard overview (risk distribution chart)
• The Metrics API (GET /v1/metrics/risk)

Strictest-action-wins

When multiple policies match the same tool call, the most restrictive action applies:

Policy A: toolName=deploy → warn
Policy B: inputPayload.env=production → require_approval
Policy C: agentId=untrusted-agent → block

Result: BLOCKED (block > require_approval > warn)

This ensures that a permissive policy can never override a stricter one.

Metrics

The risk engine feeds into the metrics system:

• Risk distribution — breakdown by severity level over time
• Top risky tools — which tools trigger the most risk events
• Agent risk profiles — which agents have the highest risk activity
• Policy match rates — how often each policy fires

Access via GET /v1/metrics/risk or the dashboard analytics page.