Glossary
These are the terms HYPR AgentPass™ and its documentation use. Each entry defines the idea, and links to the page that covers it in full.
Agents and People
| Term | Definition |
|---|---|
| Acts as | Whose access an agent uses. A local agent acts as its owner. A cloud agent acts as the person chatting with it. A workload agent acts with its own role. The Agents page filters on it: A person or Workload. See Where an Agent Runs |
| Agent | Software that uses a model and tools to complete tasks. AgentPass governs local, cloud and workload agents. See Where Agents Run |
| Agents page | The one list of every agent the gateway knows, local and cloud, with filters for provider, Acts as, Runs and agent roles. See Find an Agent |
| Cloud agent | An agent a provider hosts, such as an Anthropic managed agent, an AWS Bedrock AgentCore agent or a Copilot Studio agent. It carries a gateway-issued credential, and acts for whoever is chatting with it. A workload agent set to run in the cloud also counts as a cloud agent. See Cloud Agents |
| Evaluator | An agent role. An evaluator is a cloud agent assigned to review coding-agent sessions and return findings. See Session Watch |
| Import | Bringing agents that already exist at a provider into AgentPass, so the gateway can identify and govern them. Assistants on people's machines need no import: they appear when they first connect. See Import and Create Agents |
| Local agent | An agent that runs on a person's machine, such as Claude Code, Codex or Cursor, or a workload agent set to run on a machine. See Local Agents |
| Mobile access | Whether people can reach an agent through the HYPR mobile app. Turning it off signs out the agent's mobile app sessions. See Act on an Agent |
| Needs attention | An agent with an import conflict or an expired credential. The Agents page can show only these agents. See Read the List |
| Owner | The person accountable for an agent. A local agent's owner is the person who onboarded it. A workload agent's owner is named when it is created. See The Accountable Owner |
| Policy author | An agent role. A policy author is a cloud agent designated to draft policies. See Choosing Who Drafts Your Policies |
| Stage | Where an agent is in its life, separate from whether it is on: Production, Retain data only or Retired. Leaving Production turns the agent off, and only a Production agent can be turned on. See Lifecycle Stages |
| Status | Whether the gateway accepts an agent's requests now: Active or Off. An agent can also show Import conflict, Credential expired, or its stage. See Read the List |
| Subagent | An agent branch that performs part of a parent session's work. See Subagent Attribution |
| Supervisor | A word with three meanings. An agent's supervisor is the person accountable for the agent, who approves its actions and talks to it. The Supervisor role is a built-in role whose permissions reach only the person's own agents and records. In permission names, Supervisors means the records of people, so Supervisors: View opens the Users tab of People. See Being a Supervisor and Permissions by Role |
| Turn off | Stopping an agent at the gateway. Its next call is refused, its live sessions are stopped or interrupted, and everything about it is kept, so turning it back on restores it. See Turning Off an Agent |
| User | Any person AgentPass knows, whether they sign in to the console, supervise agents or only use them. Users are listed on the Users tab of People, as Enrolled, Invited or Offboarded. A user's role decides what they can do. See Users |
| Workload agent | An agent that runs on its own, with a named owner, its own role and credential, a connector scope and an optional expiry. It runs on a machine or in the cloud. See Workload Agents |
Roles and Access
| Term | Definition |
|---|---|
| Active role | The role a person who holds several works in, chosen under Role in the account menu. See Work in More Than One Role |
| Admin | The built-in role that holds every permission. It cannot be edited. See Permissions by Role |
| Auditor | The built-in read-only role. It reads every configuration and record, and exports guardrail evidence. See Permissions by Role |
| Model governance owner | The built-in role that registers and governs judges, and reads guardrail posture and the Request log. See Permissions by Role |
| Operator | The built-in role for day-to-day operations. It reads broadly, decides approvals, changes agent status, and manages its own policies and cloud agents. See Permissions by Role |
| Role | A named set of permissions, each at All or Own scope. A person's role decides what they can do in the console, and which tools their agents may call. The console shows a built-in role without its stored AgentPass prefix. See Roles and Permissions |
| Role claim | The sign-in token claim whose values are mapped to AgentPass roles. See Populate Roles From Your Identity Provider |
| Scope | How far a permission reaches: All records, or only the person's Own. See How a Permission Works |
| Security analyst | The built-in role for security operations. It triages guardrail alerts, and contains agents and sessions. See Permissions by Role |
Gateway, Connectors and Policies
| Term | Definition |
|---|---|
| Agent Portal | The browser page on the gateway where people chat with provider-hosted cloud agents, under the same controls as any other traffic. It is for people using agents. See Agent Portal |
| Chat channel | Two-way messaging between an agent and its supervisor, for Claude Code and Codex. See Having Conversations with Your Agent |
| Channel connector | A connector that brings people's messages from a chat platform, such as Slack, to a cloud agent. See Connectors |
| Connector | The configuration for one provider or tool the gateway reaches: an AI provider, an MCP server, a REST API, a cloud-agent platform or a channel. See Connectors |
| Console | The AgentPass administration web app. People with a role sign in to it to configure the gateway, decide approvals and review activity. Each person sees the pages their role allows. See Console Navigation |
| Endpoint Guard | The workstation client that reports AI tool activity, and applies managed routing, MCP provisioning and egress controls. See Endpoint Guard |
| Enforce mode | A policy mode that applies its decision to live traffic: allow, block or require approval. See How Policies Are Evaluated |
| Observe mode | A policy mode that records what a policy would decide, without changing the request. See How Policies Are Evaluated |
| Onboarding mode | The connector setting that decides how the gateway treats a caller it does not know yet. See Onboarding Mode |
| Policy | A condition and an action that allow, block or require approval for agent traffic. See Policies |
| Preview | The tag the console shows beside Guardrails, Conversations and Access reviews |
Supervision
| Term | Definition |
|---|---|
| Approval | A supervisor's decision on an agent action a policy held, or on an agent's first connection: approve or deny. See Approving and Denying Agent Actions |
| Signed approval | An approval carrying the supervisor's passkey signature. Every approval is signed when Require signed approvals is Required (production). A denial needs no passkey. See Decide Which Passkeys Count |
Data and Sessions
| Term | Definition |
|---|---|
| Data ceiling | The highest data classification a connector may receive. See A Connector's Data Ceiling |
| Forensics Store | Long-term storage for your organization's events, when your deployment has one connected. See Forensics Store |
| Request log | The record of every request that passes through the gateway, with the agent, the person, the connector and the decision. See Request Log |
| Session | A correlated unit of agent work that joins related requests and tool calls into one record. The gateway also tracks what a session has read, so later calls are judged on it. See Agent Session Forensics |
| Tag | A data category assigned to tools. A tag may carry a floor tier, the lowest tier anything carrying the tag can be. See Tags |
| Tier | A ranked data sensitivity level, such as Confidential. Policies compare tiers by rank. See Tiers |
Guardrails
| Term | Definition |
|---|---|
| Alert | The record a guardrail raises when it holds or stops a call, or when a tool changes after approval. Repeats of the same situation fold into one alert, and the agent's supervisor is emailed about a held or stopped call. See Guardrail Alerts |
| Binding | The choice of which promoted judge answers one category of question, injection, drift or classification, and what happens when that judge is unavailable. See Binding a Judge to a Category |
| Containment | Stopping harm in progress. Contain on an alert turns the agent off and revokes its live sessions in one action. Pausing or terminating one session contains that session alone. See Containing an Agent |
| Guardrail | A control that contains indirect prompt injection. Instead of judging how content is worded, guardrails watch what a session has read, how sensitive it was and where the session now tries to send it. See Input Guardrails |
| Judge | An optional model you register to score content, for example for signs of injection. A judge's scores are evidence: your thresholds and the breaker make the decision. See Bring-Your-Own Judges |
| Pause | Freezing one live session. Every tool call it attempts is refused, reads included, until a reviewer resumes or terminates it. Other sessions of the same agent carry on. See Pause, Resume, Terminate |
| Proving Ground | A scripted pack of injection attacks run against an isolated sandbox, showing which attacks got through with guardrails off and which were held or stopped with them on. See The Proving Ground |
| Rule-of-Two | The principle the breaker enforces: a session that has read untrusted content and sensitive data must not also reach a way out unchecked. When all three meet, the breaker holds or stops the call. See The Rule-of-Two Breaker |
| Session investigator | A continuous reviewer that reads Session Watch findings, and reaches a verdict on a whole session: continue, pause or terminate. It acts only when you set it to enforce. See The Session Investigator |
| Sink | What a tool does with data: external communications, a state change, a credential or code execution. A read is not a sink. See The Sink |
| Untrusted source | A connector whose content is not trusted input. Every connector is untrusted until someone marks it trusted, with a reason. One read from an untrusted source marks the session untrusted for the rest of its life. See Source Trust |