Architecture: The Request Path
This page follows one request through the HYPR AgentPass™ gateway, check by check. For the overall picture of agents, enforcement points, the gateway and supervisors, read How It Works first.
Where a Request Enters
Each connector has its own address on the gateway, and an agent sends its traffic to that address. Endpoint Guard points supported workstation tools at those addresses, and the Agent Portal sends its conversations to them. The connector's kind decides which checks apply:
- Model connectors, such as Anthropic, OpenAI, AWS Bedrock and Microsoft Foundry, carry inference: a conversation in, a model reply out.
- MCP and REST API connectors carry tool calls. A REST API connector presents each operation of the imported API as a tool.
- Cloud-agent connectors carry turns to an agent the provider hosts.
The Path of One Request
The gateway runs these checks in order. A check that refuses the request ends it there, and nothing reaches the provider or tool.
- Identify the caller. The gateway resolves the request to one agent and one person. An agent that is turned off is refused, whatever credential it presents. An agent the gateway does not know yet is handled by the connector's onboarding mode, for example held until its supervisor approves the first connection.
- Check the role's tool access. On MCP and REST API connectors, the agent is offered only the tools its role was granted, and a call to any other tool is refused. See Controlling MCP Tool Access by Role.
- Apply session controls. A session that was paused or terminated has its tool calls refused before any policy runs. See Pausing and Stopping a Session.
- Evaluate the request policies. Policies in observe mode record what they would have decided, and change nothing. Policies in enforce mode run in priority order, and the first one that blocks or requires approval decides. A rule that cannot be evaluated counts as a match, so a broken rule blocks rather than lets traffic through. See How Policies Are Evaluated.
- Hold for a person, where a policy requires it. The supervisor approves or denies on the HYPR mobile app or in the console. See Approving and Denying Agent Actions.
- Forward the request. Where the gateway holds the tool's credential, it attaches it now, so the agent never holds it. On AWS Bedrock and Microsoft Foundry connectors, the developer's own credential passes through, and the provider authorizes it.
- Evaluate the response policies. A rule written for the response can withhold a reply before it reaches the agent.
- Record the request. The Request log keeps the agent, the person, the connector, the decision and a redacted sample. See Request Log.
Every policy reads the attributes listed in the Policy Attribute Reference.
What Each Check Is Configured By
| Check | Configured on | Who can change it |
|---|---|---|
| Onboarding mode | The connector | Connectors: Edit (Admin) |
| Role-based tool access | The role's tool grid | Roles: Manage (Admin) |
| Session controls | Guardrails | Guardrails: Manage (Admin, Security analyst) |
| Policies | Policies | Policies: Add and Edit (Admin for any policy; Operator and Supervisor for their own) |
| Signed approvals | Settings › Gateway settings | Settings: Edit (Admin) |
| Retention | Settings › Data management | Data retention: Manage (Admin) |
Permissions by Role lists every role's grants.
Related
- How It Works: agents, enforcement points, the gateway and supervisors on one page
- Identity and Token Model: the identities a request can carry
- Security Controls Reference: each control as a verifiable claim