# Certiv > AI agent governance platform that monitors, controls, and secures autonomous AI agent operations across enterprise environments. Certiv provides real-time visibility, policy enforcement, and human-in-the-loop oversight for AI agents operating in production. It enables organizations to define security policies, approve or block agent actions, audit all agent activity, and manage agent deployments at scale. --- ## Glossary ### Agent The autonomous AI entity that performs tasks, makes decisions, and orchestrates tool calls. Agents are the "brain" doing the actual work — reasoning, planning, and executing. Agents always run within an Agentic App and their activity is tracked through sessions. Note: In casual usage, "agent" often refers to the Agentic App itself, but technically the agent is the AI component inside the app. ### Agentic App A software application that contains and runs one or more agents. Examples include Cursor, Windsurf, Claude Code, and custom-built applications using AI SDKs. The Agentic App provides the runtime environment, user interface, and integration layer, while the agent inside it does the autonomous AI work. When someone says "I'm using Cursor as my agent," Cursor is the Agentic App and the AI performing tasks within it is the agent. ### Approval A human-in-the-loop checkpoint that pauses an operation until an authorized reviewer approves or rejects the underlying agent action, tool call, or policy-triggered event. ### Decision The outcome produced when a policy evaluates an agent action or tool call. A decision determines whether the action is allowed, denied, or requires human approval, and may include additional context such as the matched rule, risk level, or audit metadata. ### Deterministic Rule A policy rule that evaluates explicit tool calls and their parameters against defined conditions. Deterministic rules operate on concrete, observable data such as tool names, argument values, or resource identifiers, providing predictable and repeatable policy enforcement. ### Endpoint A machine or device that has installed the Certiv software and is relaying data about agents back to the platform. Endpoints use token-based authentication configured through Issuers and Scopes. ### Finding An observation generated when a policy evaluates an agent action or tool call. Findings can be informational (noting activity without indicating a problem) or represent a violation (indicating a policy breach). Findings provide visibility into agent behavior and support audit, compliance, and security monitoring. ### Instance A collection of related sessions within an Agentic App. Instances group sessions that share a common context or purpose, such as a project workspace or user environment. Instances are a higher-level organizational unit but are less commonly used than sessions for day-to-day tracking. ### Intent-Based Rule A policy rule expressed in natural language that evaluates the intent or purpose behind an agent's actions rather than specific tool calls. Intent-based rules use semantic understanding to assess whether agent behavior aligns with or violates policy goals, providing flexibility for scenarios where explicit tool matching is insufficient. ### Issuer A cryptographic signing key configuration used to sign User Linking Tokens (ULTs) for endpoint authentication. Issuers can be "Hosted" (Certiv-managed, where Certiv stores and manages the private key) or "BYOK" (Bring Your Own Key, where the organization manages the private key externally). ### Member A user who belongs to an organization. Members have roles and permissions that determine their access to resources and capabilities within the organization. ### Organization A top-level entity in the Certiv platform that provides a collaborative workspace for teams and enterprises. Organizations manage members, resources (e.g., Agent Control Groups, agents, tools), and governance policies. ### Out-of-the-Box Policy A default policy provided by Certiv that addresses common security and compliance requirements. Out-of-the-box policies offer ready-to-use governance rules that organizations can apply immediately or use as templates for customization. ### Policy A set of rules applied to agents and their tool calls that governs what actions are permitted, denied, or require approval. Policies define the security and compliance boundaries for agent operations, specifying conditions under which tool calls can execute automatically versus when human oversight is required. ### Profile A named configuration that organizes and groups related policies together. Profiles provide a convenient way to apply consistent governance rules across multiple agents or Agent Control Groups, enabling reusable policy bundles that can be assigned as a unit. ### Scope A permission template (also known as ULT Configuration) that defines default roles, groups, and permissions assigned to tokens generated using a specific issuer. Scopes provide granularity in controlling what access tokens grant to endpoints. ### Scout The Certiv endpoint software that runs on machines and devices to monitor agent activity. Scout relays agent data back to the Certiv platform, enabling visibility, policy enforcement, and governance across distributed endpoints. ### Session The fundamental unit of agent activity in the Certiv platform. A session represents a discrete period of agent operation, typically corresponding to a single conversation, task execution, or workflow run within an Agentic App. Sessions capture all tool calls, decisions, and interactions that occur during that period. Sessions are the primary building block for aggregating metrics, auditing activity, and organizing agent work into higher-level groupings like instances. ### STEAR Secure Task Execution and Agent Runtime — the platform execution layer that supplies isolated infrastructure, governance, and policy enforcement for agent operations. ### Agent Control Group A logical container that organizes endpoints, defines operational policies, and controls access to resources within the Certiv platform. Agent Control Groups provide isolation boundaries and governance structures for endpoint operations. Their members are endpoints (`endpoint_count`); an endpoint with no explicit control group is labeled Unassigned. Agents are identities with no group membership. ### STEAR Worker An isolated execution unit within the STEAR system that runs agent workloads with dedicated resources, security controls, and health monitoring. ### Tool A discrete capability or service that agents invoke to perform work, such as making API calls, processing data, or interacting with external systems. ### Tool Call An individual interaction where an agent requests a tool to perform an operation, progressing through queueing, execution, and completion stages that can be visualized or replayed. ### Tracked Token A User Linking Token (ULT) whose lifecycle is monitored by the platform. Tracked tokens have observable states: Awaiting (generated but not yet used for authentication), Activated (successfully used to authenticate an endpoint), or Revoked (invalidated and no longer valid for authentication). ### User Linking Token (ULT) A cryptographically signed token used to authenticate endpoints with the Certiv platform. ULTs are generated using an Issuer's signing key and carry identity and permission information defined by a Scope. Their lifecycle can be monitored as Tracked Tokens, progressing through Awaiting, Activated, or Revoked states. ### Violation A finding that indicates an agent action or tool call breached a policy rule. Also referred to as a Policy Violation. Violations represent non-compliance with defined governance rules and typically trigger alerts, require remediation, or are logged for audit purposes. --- ## Agents An **Agent** is an autonomous AI entity that operates within the Certiv platform to perform tasks and execute workflows. Agents communicate with tools through the platform's visualization and management system, making decisions and orchestrating operations based on their configuration and assigned objectives. ### Purpose Agents serve as the primary execution units in the Certiv platform, providing: - Autonomous task execution and decision-making - Tool orchestration and workflow management - Real-time interaction with available platform tools - Controlled AI operations within organizational boundaries - Auditability and monitoring of AI behaviors ### Key Properties - **ID**: Unique identifier for the agent - **Name**: Human-readable identifier - **Type**: Classification of the agent's purpose or specialization - **Description**: Optional explanation of the agent's role - **Status**: Current operational state (Active, Inactive, or Error) - **Configuration**: Behavioral settings and parameters - **Created By**: User who created the agent ### Agent Lifecycle 1. **Creation**: Agent is created by authorized users, initial configuration is set, status set to active by default 2. **Operation**: Agent executes tasks and workflows, communicates with tools through the visualization system, actions are logged and audited 3. **Management**: Status can be changed (active/inactive), configuration can be updated 4. **Decommissioning**: Agent can be removed from the system, historical data and logs are preserved, associated resources are cleaned up ### Relationships - **Tools**: Agents interact with tools to execute tasks. Tool calls are visualized and can be replayed. - **Users**: Created and managed by platform users. Agent actions are attributed to creating users. - **Organizations**: Agents operate within organizational boundaries. Organizational policies apply to agent behavior. --- ## Policies A **Policy** is a rule-based governance mechanism that controls and regulates agent behavior within the Certiv platform. Policies define automated decision-making criteria that determine whether agent operations should be allowed, blocked, require human approval, or trigger specific workflows. ### Purpose - **Automated Governance**: Rule-based decision making for agent operations - **Risk Management**: Preventing dangerous or unintended agent behaviors - **Compliance**: Ensuring operations meet regulatory and organizational requirements - **Security Enforcement**: Protecting systems and data from unauthorized access - **Human Oversight**: Determining when human intervention is required ### Policy Decision Types - **Allow**: Operation proceeds without restriction - **Block**: Operation is immediately terminated - **Pause**: Operation is temporarily suspended, requires human approval to continue - **Block Gracefully**: Operation is terminated with clean shutdown and detailed explanation - **Human Approval**: Operation requires explicit human authorization, queued for approval workflow ### Policy Rule Types **Simple (Deterministic) Rules**: Condition-based if-then logic with fast, deterministic evaluation. Supports operators including equals, not_equals, contains, not_contains, greater_than, less_than, in, not_in, and regex. Conditions can evaluate tool_name, tool_type, agent_id, user_id, time_of_day, request_size, and custom criteria. **LLM-Based (Intent) Rules**: Natural language policies with contextual understanding. Provides flexible, adaptive decision making through intent recognition and dynamic evaluation. ### Policy Scope Policies can be applied at multiple levels: - **Platform-Wide**: All operations across the platform - **Organization-Level**: All operations within an organization - **Agent Control Group-Level**: Specific groups and their endpoints - **Agent-Specific**: Individual agents - **Tool-Specific**: Specific tool types or instances --- ## Tools A **Tool** is a functional capability that agents can access and utilize to perform specific operations. Tools represent discrete services, APIs, or capabilities that extend agent functionality. ### Tool Categories - **Codebase Retrieval**: Code searching, file navigation, version control integration - **File Editor**: File creation, modification, deletion, format conversion - **Web Search**: Internet search, web scraping, API integration - **Database**: Database queries, data modification, schema management - **API Client**: HTTP/REST communication, authentication handling - **Git Operations**: Version control (commit, push, pull), branch management - **Testing**: Test execution, validation, performance testing - **Deployment**: Application deployment, infrastructure provisioning - **Monitoring**: System monitoring, metrics collection, alerting - **Security**: Vulnerability assessment, access control, encryption - **ML Inference**: Machine learning model execution, NLP, computer vision - **Custom**: Organization-specific tools and integrations ### Tool Integration Tools integrate with the platform through the Model Context Protocol (MCP): - **Server Components**: MCP servers provide tool capabilities - **Capability Discovery**: Dynamic tool capability enumeration - **Request/Response Handling**: Structured message exchange --- ## Agent Control Groups A **Agent Control Group** is a logical container and management unit that organizes endpoints, defines operational policies, and controls access to resources. Its members are endpoints; an endpoint with no explicit control group is labeled Unassigned. Agents are identities with no group membership. ### Endpoint Membership - Endpoints (machines and devices running Scout) are assigned to a group and inherit its policies. - A group's contents and counts (`endpoint_count`) refer to its endpoints, never to agents. - An endpoint may be assigned to a single group; one with no explicit control group is Unassigned. ### Key Properties - **Endpoint Count**: Number of endpoints assigned to the group (`endpoint_count`) - **Owner Type**: Either user or organization - **Messaging**: Dedicated messaging for group communications and job processing - **Deployment Type**: Type of deployment environment ### Ownership Models - **User Ownership**: Individual users own and manage the group. Suitable for personal projects. - **Organization Ownership**: Organization-level management. Supports enterprise governance with role-based access. --- ## Organizations An **Organization** is a top-level entity providing a collaborative workspace following a GitHub-style model. ### Member Roles - **Owner**: Full administrative control, can delete organization, manage billing - **Admin**: Near-full capabilities, manage members and settings (except billing/deletion) - **Manager**: Manage Agent Control Groups, approve operations, view analytics - **Member**: Standard access, can be assigned to groups, participate in approvals - **Viewer**: Read-only access, suitable for external stakeholders or auditors ### Invitation Workflow Organizations use invite-only membership: invitations are created with a specific role, sent via email, and can be in states of Pending, Accepted, Rejected, Withdrawn, or Expired. ### Key Settings - Members Can Create Agent Control Groups (configurable) - Require 2FA for all members - Max Agent Control Groups and Max Members limits - Public or Private visibility --- ## Environments An **Environment** is a logical and technical boundary defining the operational context for agents, tools, and workflows. ### Environment Types - **Development**: Relaxed controls, test data, broad developer access - **Staging**: Production-like controls, sanitized data, approval-based deployments - **Production**: Full security controls, live data, highly restricted access, full SLA - **STEAR**: Specialized for STEAR workers, with enhanced workload isolation --- ## Approvals An **Approval** is a human-in-the-loop workflow providing oversight for agent operations. ### Approval Types - **Tool Call Approvals**: High-risk operations, sensitive data access, system changes - **Policy-Triggered Approvals**: Policy violations, exception requests, compliance reviews - **Administrative Approvals**: User management, organization changes, security operations ### Approval States Pending → Assigned → In Review → Approved/Rejected/Withdrawn/Expired/Escalated ### Workflows Standard flow: Request Generation → Assignment → Notification → Review → Decision → Notification → Execution → Documentation Supports time-based escalation, authority-based escalation, approval delegation, and bulk operations. --- ## Audit Logs **Audit Logs** are comprehensive, tamper-evident records of all significant activities within the platform. ### Event Categories - **User Activity**: Authentication, authorization, profile changes, session management - **Agent Operations**: Lifecycle events, task execution, tool interactions, policy evaluation - **Agent Control Group Management**: Group operations, endpoint assignments - **Organization Activities**: Management, invitations, role changes, settings changes - **Policy and Approval Events**: Enforcement, approval workflows, rule changes - **Security Events**: Access control, authentication failures, suspicious activity - **System Operations**: Configuration changes, maintenance, performance events ### Integrity Audit logs use cryptographic hash chains, digital signatures, and Merkle trees for tamper detection. Logs are stored in immutable, encrypted, geographically distributed storage. ### Integration - SIEM platforms for real-time streaming and alert generation - GRC platforms for governance, risk, and compliance - Business intelligence for analytics and reporting --- ## STEAR Workers A **STEAR Worker** is a specialized execution unit providing isolated, secure computing resources for agent operations. ### Worker Types - **Compute Workers**: CPU-intensive processing, data analysis, algorithmic operations - **Storage Workers**: File operations, database interactions, data archival - **Network Workers**: API calls, web scraping, external service integration - **Specialized Workers**: ML inference, security scanning, custom business logic ### Isolation STEAR Workers run with configurable isolation matched to the sensitivity of each workload. ### Integration - Task queues for distribution and management - Container orchestration - Health monitoring via heartbeat system - Auto-scaling based on workload demands --- ## Architecture Overview ### Components - **API Server**: REST API for governance, policy enforcement, and audit - **Web Application**: single-page web application - **Python SDK**: Agent monitoring, policy enforcement, tool call tracking - **Scout**: Endpoint monitoring software for distributed agent visibility - **STEAR Runtime**: Isolated execution layer with workers and sidecars ### Authentication - JWT-based authentication with refresh tokens - HMAC authentication for agent instances - User Linking Tokens (ULTs) for endpoint authentication - Support for SSO (SAML, OAuth) integration ### Data Flow - Asynchronous messaging and event streaming for agent activity - Caching for low-latency reads and session state - Durable persistent storage for records and audit history - Search and logging over agent and audit data - Metrics-based monitoring and alerting