Key Concepts

A quick reference for the terms you'll see throughout the Behavry dashboard.

Agent

An AI system (Claude, ChatGPT, a custom LangChain bot, etc.) that has been enrolled in Behavry. Each agent has a unique identity, a behavioral baseline, and a risk tier.

Session

A single working period for an agent. Sessions start when an agent connects and end when it disconnects or times out. All activity within a session is grouped for audit and analysis.

Tool Call

A single action an agent takes — reading a file, querying a database, sending a message, calling an API. Each tool call is authenticated, evaluated against policy, and logged.

Policy

A rule (or set of rules) that governs what agents are allowed to do. Policies are evaluated for every tool call. They can allow, deny, escalate to a human, or restrict the agent to a limited mode.

Escalation

When a policy requires human review before an action proceeds. The agent pauses, a notification is sent to your security team, and the action proceeds (or is blocked) based on the reviewer's decision.

Risk Score

A continuously updated score (0–100) reflecting how risky an agent's recent behavior appears. Behavry considers tool variety, action sensitivity, anomaly signals, and known threat patterns.

Enrollment Token

A single-use token used to register a new agent with Behavry. After registration, the agent receives a persistent JWT credential for all future authentication.

Baseline

A statistical model of normal behavior for an agent, built from its historical activity. Deviations from baseline trigger anomaly alerts.

DLP Pattern

A data-loss prevention rule that detects sensitive content (PII, credentials, financial data, etc.) in agent inputs and outputs.

Tenant

A single organization's isolated Behavry environment. All agents, policies, audit logs, and settings belong to a tenant.