What is behavioral monitoring for AI agents?

Behavioral monitoring observes an AI agent's actions over time and compares them to an established baseline. When the agent's behavior deviates significantly from what is expected, the system raises an alert. This catches problems that per-action policy rules miss.

Why per-action rules are not enough

A policy engine evaluates each tool call independently. It asks: "Is this specific action allowed?" But some threats only become visible when you look at patterns:

• A single database query is fine. A hundred queries in ten seconds is a data exfiltration attempt.
• One denied action is normal. Twenty denials in a row means the agent is trying to do something it should not.
• Using a search tool is expected. Suddenly switching from search to email tools suggests the agent's goal has changed.

Behavioral monitoring answers a different question: "Is this agent acting like it normally does?"

What gets monitored

A behavioral monitor tracks metrics over sliding time windows:

• Action rate: Tool calls per minute
• Denial rate: Percentage of actions blocked by policy
• Tool distribution: Which tools are called and how often
• Error rate: How often tool calls fail
• Argument entropy: Statistical variation in tool arguments
• Session duration: How long the agent has been active

Baseline and deviation

During normal operation, the monitor builds a baseline: the typical values for each metric. Once the baseline is established, the monitor flags deviations.

For example, if an agent normally calls tools 5 times per minute and suddenly starts calling 50 times per minute, the rate is 10x the baseline. The monitor flags this as an anomaly.

Detection algorithms

Two common algorithms for anomaly detection:

EWMA (Exponentially Weighted Moving Average): Maintains a running average that gives more weight to recent values. Alerts when the current value deviates from the average by more than N standard deviations. Good for detecting gradual drift.

CUSUM (Cumulative Sum): Accumulates small deviations from the expected value. Alerts when the accumulated deviation exceeds a threshold. Good for detecting subtle, persistent changes.

Alert response

When the monitor detects an anomaly, it can:

• Log the alert for later review
• Send a notification to the operations team
• Tighten the active policy (switch to a stricter ruleset)
• Terminate the agent session

The appropriate response depends on the severity and your risk tolerance.

Sentinel

Authensor's behavioral monitor is called Sentinel. It runs in-process with zero dependencies, tracking metrics per session. It supports both EWMA and CUSUM detection with configurable thresholds and alert handlers.