Risk Scoring

A daily composite score per agent, used to prioritize review and gate sensitive changes.

Risk scoring computes a single 0.0-1.0 score per agent every day, combining six signals about that agent's stability, ownership, and recent behavior. Use it to decide which agents need attention first — and which need a second pair of eyes before deployment.

How the score is computed

A daily worker runs at 02:00 UTC, evaluates every approved agent, and writes:

MusterAgent.riskScore — the raw decimal
MusterAgent.riskLevel — bucketed level (see below)
MusterAgent.riskFactors — JSON breakdown of which signals contributed
MusterRiskScoreHistory — daily snapshot for trending

The six signals (default weights)

Signal	Default weight	When it adds risk
No registered owner	+0.15	Agent has no `team` / `department` set
Status not yet `APPROVED`	+0.20	Agent is `DISCOVERED` or `PENDING_REVIEW`
Unknown framework	+0.15	`framework = "UNKNOWN"`
Active anomalies	+0.10 each	Capped at +0.25 (3+ anomalies)
Pass rate below 70%	+0.15	7-day rolling pass rate on Muster scores
Missing description	+0.10	`description` is null or empty

The signals sum, clamp to [0.0, 1.0], and bucket:

Level	Score range
`LOW`	0.0 - 0.3
`MEDIUM`	0.3 - 0.5
`HIGH`	0.5 - 0.7
`CRITICAL`	0.7+

Weights are tunable per project via MusterProjectTuning. The Auto-Instrumentation flow can suggest weight changes based on your fleet.

Where you see it

Agent inventory — every row shows the level (color-coded badge) and the raw score on hover.
Risk dashboard — project-wide distribution by level plus the movers from yesterday's snapshot.
Agent detail — full riskFactors breakdown so you can see exactly which signals are triggering.

How risk gates approval workflows

Approval policies (covered in detail in the upcoming Governance guide) can require additional sign-off based on risk level. The most common pattern: any transition into APPROVED for an agent currently rated HIGH or CRITICAL requires a second approver from a designated group.

This is why the lifecycle and the score are two sides of the same coin: the score tells you which agents need scrutiny, the workflow ensures scrutiny happens before they go live.

Reducing risk

The risk score is read-only — you don't change it directly. To lower an agent's score:

Assign a team and department
Move from PENDING_REVIEW to APPROVED
Set the framework field to a real value
Resolve open anomalies
Improve pass rate (fix the underlying agent, not the metric)
Add a description

Each fix removes its corresponding signal from the next daily run.