AWS Bedrock AgentCore Agents
Overview
Amazon Bedrock AgentCore is where teams run their own agents, as containerized agent code on the AgentCore Runtime or as managed harnesses. An AgentCore connector brings those agents into HYPR AgentPass™. The gateway imports them, lets your people chat with them, runs them on schedules and hands them governed gateway tools. Their actions are then subject to your policies and audit trail.
AgentPass never deploys or modifies your agent's code. The agents stay yours, in your AWS account. AgentPass invokes and governs them.
Before You Begin
Choose how the connector authenticates to AWS:
- IAM role (recommended). Create a role in the AWS account that hosts your agents. In the connector form, enter its ARN, and copy the generated trust policy into the role's trust relationship. The gateway assumes the role with its deployment identity, so AgentPass stores no long-lived credential.
- This deployment's AWS identity. Use this option when AgentCore and the gateway run in the same AWS account. Grant the gateway's deployment identity the baseline policy. No connector credential is stored.
- Access key (advanced). Create an IAM user in the AWS account that hosts your agents, and store its access key in the connector. AgentPass encrypts the secret and never shows it again after saving.
Attach this baseline policy to the role, deployment identity or IAM user that the connector uses:
{
"Version": "2012-10-17",
"Statement": [
{
"Sid": "HyprAgentCoreIngestAndInvoke",
"Effect": "Allow",
"Action": [
"bedrock-agentcore:ListAgentRuntimes",
"bedrock-agentcore:GetAgentRuntime",
"bedrock-agentcore:ListAgentRuntimeEndpoints",
"bedrock-agentcore:InvokeAgentRuntime",
"bedrock-agentcore:StopRuntimeSession",
"bedrock-agentcore:ListHarnesses",
"bedrock-agentcore:GetHarness",
"bedrock-agentcore:UpdateHarness",
"bedrock-agentcore:InvokeHarness"
],
"Resource": "*"
}
]
}
The UpdateHarness action lets AgentPass attach and rotate gateway-tool credentials on a harness. If you do not use direct harness attachment, you can omit it.
The StopRuntimeSession action lets AgentPass stop an agent's running sessions in AWS when you turn the agent off. Without it, the agent is still turned off and the gateway refuses its calls, but sessions already running in AWS continue until AWS ends them. A harness has no session to stop this way, so for a harness agent the gateway's refusal is what stops it.
The connector form shows the same actions on its IAM permission policy card. Click Copy on the card to copy the policy.
Create the Connector and Import Your Agents
- In the console, open Connectors.
- Click Add Connector.
- Under Cloud Agents, choose AWS Bedrock AgentCore.
- Under How should the gateway authenticate to AWS?, choose a mode.
- Complete the connector details described in the following table.
- Optional: Under Connection, click Test connection. The test uses the selected mode and what is in the form, so you do not need to save first.
- Click Create Connector. AgentPass lists the account's agent runtimes and harnesses right away, imports each as an agent in Cloud Agents and opens the connector's Agents tab. A message reports how many agents were imported.
The connector is created even when the first import does not complete. In that case, click Sync agents from AWS on the connector to import the agents.
The connector form has these fields.
| Field | What to enter |
|---|---|
| Account label | A friendly name for the AWS account, shown in the console |
| How should the gateway authenticate to AWS? | IAM role (the default) for a cross-account role, This deployment's AWS identity for same-account AgentCore, or Access key for an IAM-user key. The first two are unavailable when the deployment has no AWS identity to assume from |
| Role ARN | For IAM role, the ARN of the role in your AWS account that trusts the gateway. Copy the generated trust policy into that role's trust relationship |
| AWS Access Key ID and AWS Secret Access Key | For Access key, the IAM user's key. The secret is stored encrypted and never shown again after saving |
| Region | The region your AgentCore runtimes live in |
| Owner email | The supervisor who owns agents imported from this account. This person approves until agents are assigned to people |
Each imported agent appears once, named from AWS, with Synced from provider on its Overview tab. Re-sync the connector whenever you add agents in AWS. Removing the connector removes its imported agents from AgentPass.
The header switches, Evaluator, Policy author and Enabled, work as they do for any provider-managed agent. See Provider-Managed Cloud Agents.
The Harness configuration section is particular to AgentCore. It is stored in AgentCore itself, so a change saved here writes through to AWS, and the AWS console shows the same values.
The agent is built and owned in your AWS account, so AgentPass does not edit its model or code. There is no Edit configuration for an AgentCore agent. The one exception is a harness's System prompt and Memory in the Harness configuration section, which Save to AWS writes through to AWS. Saving needs permission to synchronize agents. AgentPass manages everything around it: who may use it, which gateway tools it holds, when it runs and whether it is enabled.
Talking to an AgentCore Agent
Open the agent, and click the Chat tab. AgentPass accepts your message immediately, and the agent's reply appears in the transcript when its turn completes. AgentCore turns can take a while for complex work. A spinner shows in the transcript while the agent works, and the conversation view updates itself when the reply arrives.
AgentPass stores the transcript. The conversation history is in your governance domain and visible in the console, even though the agent ran in AWS.
Hooking In Gateway Tools
An AgentCore agent's own tools are built into its code in AWS. AgentPass adds governed MCP tools: gateway connectors the agent calls with a credential AgentPass mints for it. Every call is then policy-checked, approvable and audited.
- Open the agent's Tools tab.
- Click Configure tools. The Connect to gateway tools drawer opens.
- Select the MCP server and REST API connectors this agent may call. Your selection bounds the credential, so the agent can never call a connector outside it.
- Mint the credential. AgentPass issues the agent a token for those connectors and shows it once, so copy it immediately. You can rotate it here at any time, which invalidates the old one.
- Give the tools to the agent. How depends on the kind of agent, as described in the following table.
| Agent kind | How the tools are wired |
|---|---|
| Harness | Click Attach selected to harness. AgentPass configures the selected gateway tools directly on the harness in AWS, with no copy and paste |
| Runtime | Copy the generated configuration into the agent's AgentCore Identity in AWS, so your agent code picks up the tools |
The generated configuration has one entry per allowed connector:
{
"mcpServers": {
"your-connector": {
"url": "https://<your-gateway-url>/your-connector/mcp",
"headers": { "Authorization": "Bearer <the minted token>" }
}
}
}
From then on, the agent's calls to those tools go through the gateway like any other governed traffic.
Per-User Credential Isolation
An AgentCore agent is one shared runtime that many people may invoke. By default, all its tool calls run as a single shared identity, whoever is chatting with it.
Per-user credential isolation changes that. With it on, each tool call runs as the person using the agent, with their connected accounts, consents and approvals. Nothing changes inside your agent's code, and you create no per-user resources in AWS.
- The agent's header shows its mode with a Per-user isolation or Shared identity tag.
- You turn isolation on or off per agent in the Tools drawer. If the toggle is unavailable in your deployment, confirm with your HYPR representative.
- With isolation on, the gateway refuses a call whose user it cannot determine, instead of running it as anyone else.
When you assign a shared-identity agent to a group or to more than one person, AgentPass asks you to acknowledge that its tool calls are not isolated per user. The audit trail records the acknowledgement.
Watching the Agent's Model Usage
The agent's tool calls are governed through the gateway. You can also make its model inference visible in the Request Log:
- Create an AI provider connector in Observe only onboarding mode for the model provider the agent uses.
- Have your agent send its inference through that connector, with its AgentPass agent identifier in a request header. The Tools drawer shows the exact header value to copy.
The Request Log attributes inference observed this way to the agent. While the agent is turned off, the gateway refuses the inference it attributes to the agent.
Unlike Anthropic's platform, AgentCore enforces no spend limit of its own. The gateway's limits are the only ceiling on what an AgentCore agent can consume. Keep that in mind before you schedule one aggressively or tag one as a Session Watch evaluator.
Example: Governing a Support Harness
Suppose your team runs a support-triage harness on AgentCore. You want it to file tickets under governance, instead of only drafting answers.
- Import it. Add the AgentCore connector for the AWS account, and sync. The harness appears in Cloud Agents.
- Give it a tool. In its Tools drawer, select your ticketing REST API connector, and mint the credential. It is a harness, so click Attach selected to harness. No code change is needed.
- Add a guardrail. Create a policy that requires approval for the ticket-creation tool, so a supervisor confirms each ticket the agent wants to file.
- Try it. On the Chat tab, paste a support email and ask the agent to triage it. When it calls the ticketing tool, the approval request reaches the supervisor's phone. Approve it, and the call completes. The whole exchange then appears on the Activity tab and in the Request Log.
- Schedule it. Add an automation that runs it every hour against the unassigned queue. See Automating Cloud Agents.
Related
- Managed Cloud Agents: Overview — the shared console, lifecycle and governance
- Automating Cloud Agents — schedules and run history. AgentCore agents can be scheduled; webhook triggers are Anthropic-only
- Connecting AWS Bedrock — the separate connector for routing model inference to Bedrock