Policies and Guardrails
HYPR AgentPass™ decides what every agent may do at the gateway, as each request passes. This section covers each control, what it decides, and who can set it.
The Controls
| Control | What it decides | Who sets it, out of the box | Start with |
|---|---|---|---|
| Identity | Which agent sent a request, and which person is accountable for it | Admin and Operator import agents and turn them off and on. A supervisor approves their own agents' onboarding | Identity |
| Policies | Whether a request is allowed, blocked or held for approval | Admin, and Operator and Supervisor for their own policies. Every role except Model governance owner can read them | Policies |
| Tool access | Which MCP tools an agent may see and call, by its supervisor's roles | Admin | Controlling MCP Tool Access by Role |
| Data classification | How sensitive the data each tool and connector handles is | Admin | Data Classification |
| Session tracking | What a session has already read, carried forward to its later calls | Admin | Session Tracking |
| Guardrails | Holding or stopping a call that looks like indirect prompt injection, alerting on it, and containing the agent | Admin and Security analyst work alerts. Admin saves settings | Guardrails |
| Session control | Freezing one live session, and resuming or ending it | Admin and Security analyst, and the session investigator | Pausing and Stopping a Session |
| Session investigator | Whether a flagged coding-agent session should continue, pause or terminate | Admin | The Session Investigator |
Each page names the permission a task needs. Permissions by Role lists every role's permissions by area.
Tools Governance
AgentPass governs tools on four axes. Each axis answers a different question, and has its own page.
| Axis | Question it answers | Where it is set |
|---|---|---|
| Role-based access (RBAC) | Which tools may an agent call? The tools follow the supervisor's roles, both in the list an agent sees and at call time | Controlling MCP Tool Access by Role |
| Scope | Which connectors does this rule or credential reach? A policy is scoped to connectors, and a workload agent carries a named set of connectors it may use | How Policies Are Evaluated · Workload Agents |
| Lineage | Which thread of which session made this call? Sessions group a conversation's requests, subagent attribution splits them, and an evaluator can review them | Subagent Attribution · Session Watch |
| Classification | What does this tool handle? A tier and tags on each tool and connector let a policy target the sensitivity of the data, not the name of the tool | Data Classification |
Session tracking extends classification across calls. It carries what a session has already read forward to its later calls. See Session Tracking.
Related
- Session Watch, The Session Investigator and Pausing and Stopping a Session — reviewing a live session and, when needed, freezing it
- Guardrails — defending agents against indirect prompt injection
- The Four Controls — where tools governance sits among the gateway's controls
- Roles and Permissions — the roles the RBAC axis reads