Risk Scores & Priorities

AI Agent Watch scores risk at two levels: a risk score on each event, and a risk score for the whole session once it ends. Both exist so you can either react to a single risky action as it happens, or judge a session as a whole after the fact.

Priority

Priority (Info, Medium, High, or Critical) is set per alert rule - you choose it when you create or edit a rule. It is not calculated from anything; it's a judgment call you make about how serious a given condition is.

Event risk score

Every event gets a risk score, derived directly from the priority of whichever alert rule matched it (the highest one, if more than one matched):

Priority	Risk score
Info / no rule matched	0
Medium	10
High	20
Critical	30

An event that matches no rule stays at Info priority with a risk score of 0. This score is the raw input to the session risk score below.

Session risk score

When a session ends (agent_session_end), AI Agent Watch combines the risk scores of every event in that session into a single session risk score from 0 to 100.

Rather than simply adding scores together - which would let a long session with many small findings look artificially critical - each additional risky event pushes the score closer to 100 with a diminishing effect the closer it already is to the ceiling. In practice that means:

• One or two Critical-priority events in a session will push it well up the scale
• A session made up mostly of low-risk events climbs more slowly and rarely reaches the top of the range on volume alone
• The score can approach but not exceed 100

Where these scores show up

• Event risk score and priority are shown on every event and used to sort/filter activity.
• Session risk score is shown on the session summary and used in reports that break sessions down by risk band: Clean (0-30), Medium (31-60), High (61-80), and Critical (81-100). See the Agent Sessions report in Reports.
• The default "Critical-Risk Session" alert rule fires when a session's risk score crosses a high threshold, so you can be notified about a session as a whole rather than only reacting event-by-event. You can also write your own rules against sessionRiskScore on agent_session_end events.