Manage Agent
Create an Agent
Select creation method
After you click Create an Agent, select a creation method:
Managed Harness: The Agent runs in managed mode after you select a Harness. The Agent template is optional. Available Harness types are QwenPaw, Qoder, Codex, Pi, and DeepSeek Harness. For the capabilities of and differences between these Harnesses, see Managed Harness Agent user guide. The following steps use this method as an example.
Custom code/Image: Deploy the Agent with a container image and customize the logic. For more information, see Deploy and use a high-code Agent.
Quick creation: Form-based low-code creation. Coming Soon.
Basic information
Sign in to the AgentCore console.
In the left navigation pane, click Agent.
Click Create an Agent.
Select Managed Harness.
Select a Harness.
Choose whether to use a template to configure the Agent.
Enter the Agent name.
Connect a model and select a model.
In the Reasoning effort drop-down list, select a level. Valid values: default, off, low, medium, high, xhigh, and max. The default value is default.
(Optional) Enter a Description.
Token quota
You can enable or disable the token quota management feature by using the switch, enter the quota limit, and select the metering cycle (day, week, or month).
File configuration
Enter the system prompt for the Agent to describe its role, responsibilities, and behavior.
Prompt variables
Define variables and reference them in the System prompt to replace parts of the prompt in different invocation scenarios for the same Agent.
In the upper-right corner of System prompt, click Prompt variables (N) to open the dialog box. N indicates the number of variables that are currently defined.
Click Add variable, and then enter the Variable name, Description, and Default value. A variable name must start with a letter or an underscore and can contain only letters, digits, and underscores. A blank description or default value is saved as an empty string
"".In the System prompt, enter
/or{to insert a variable, or reference a variable directly as{variable_name}. The References in prompt column in the dialog box indicates whether the variable is referenced in the system prompt.Click Save to update both the system prompt content and the variable list. When you invoke the Agent, variables that are not provided use their default values. A blank default value is replaced with an empty string.
Import and export variable definitions in bulk
In the Prompt variables dialog box, use Advanced operations · Import / Export JSON to import or export variable definitions as JSON. The JSON structure is an array, and each element contains the name, description, and default_value fields.
To import definitions, click Apply to form. Importing replaces only the variable definitions, does not automatically modify references in the prompt, and overwrites the current draft. Click Copy JSON to export the variable definitions. After you make changes, click Save for the changes to take effect.
If no Agent template is selected, you only need to enter the system prompt. SOUL.md and AGENTS.md are not required.
Agent asset configuration
The creation page provides five asset tabs: Skill, SubAgent, Tools, Knowledge base, and Credential. Bind assets as needed.
Skill
You can select available Skills from the left side to bind them to the Agent.
SubAgent
Click Add SubAgent to add a SubAgent to the Agent.
Tools
You can bind tools to the Agent on this tab.
Knowledge base
You can bind knowledge bases to the Agent on this tab. Knowledge bases are connected through MCP and are also shown under the Tools tab.
Credential
You can select available credentials from the left side to bind them to the Agent.
Resource configuration
Specifications: Select the CPU and memory required by the instance.
Set the maximum number of sessions that a single instance can host. Sessions beyond this limit are hosted by new instances.
Session TTL (seconds): Set the lifetime of a platform session. This is not the timeout of a single model generation.
Automatic scaling: When you enable automatic scaling, set the Minimum number of instances and Maximum number of instances.
Network configuration
Allow the default network interface to access the internet: Enable this option if the Agent needs to access public model services or other public endpoints.
Allow VPC access: Enable this option if the Agent needs to access the VPC resources of the workspace. When enabled, the VPC is the same as the VPC of the current workspace.
After you complete the preceding configurations, click Create Now to create the Agent.
API access
On the API access page of the Agent details page, you can invoke the Agent from external systems through a standard AG-UI endpoint and configure the inbound access security policy. The page displays the endpoint status, name, and type (public or VPC), as well as the access address generated by the server.
The API access page provides a master switch for API access.
Inbound security level
Changes to the security level or user scope take effect after you save them. The inbound security level has three tiers:
-
No authentication: Any caller that obtains the endpoint can access the Agent. This mode allows any caller with the endpoint to access the Agent. Use it only in trusted network environments.
-
Fixed API Key: The caller passes a fixed key in the specified header. The page displays the header name, key fingerprint, and key source in masked form, and supports resetting the key and obtaining authentication information.
-
User identity: Uses user JWT authentication and requires you to configure the accessible user scope. All users: Allows access by all authenticated users. Specified user list: Allows access only by the selected users. You can search for and add users on the page.
The Inbound security level area shows the currently saved access permission and authentication method, such as Fixed API Key.
Existing Agents use Fixed API Key authentication by default.
The API Key area shows the Header name, Key fingerprint, and Key source. API Keys are masked by default. Click Get Authentication Info to view authentication information.
Invocation
-
Requests use the AG-UI protocol, and the response type is
text/event-stream. For session affinity,X-AgentCore-Session-IDmust matchthreadIdin the request body. When header isolation is used, the request header name is the session header name configured for the Agent, and the value is still the session ID. -
For public network access, the page provides the AG-UI invocation address (POST · SSE) and the stop address, both of which can be copied:
-
Invocation address:
https://<access domain>/agents/<AgentID>/<Endpoint name>/ag-ui/agent -
Stop address:
https://<access domain>/agents/<AgentID>/<Endpoint name>/ag-ui/agent/stop
-
-
When User identity is selected, requests must include
Authorization: Bearer <JWT>. When Fixed API Key is selected, requests must include the API Key header specified on the page.
Authentication quick reference: When Fixed API Key is used, requests must include the API Key header.
Obtain authentication information before viewing the curl command.
Application access
On the Application access page under Identity and permissions in the Workspace, you can create and manage access applications. The list shows the application name, user source, and callback URL, and supports viewing details, editing, rotating keys, and deletion.
In the Application access area on the API access page of an Agent, you can select or add access applications for the Agent and view the application name, application ID (Client ID), user source, callback URL, and scope.
Agent details page
Access Agent details
-
Log on to the AgentCore console, and in the left-side navigation pane, choose AGENT construction > Agent.
-
Find and view an Agent on the Agent list page.
-
The top of the page shows three statistics cards: Total Agents, Running Agents, and Abnormal Agents.
-
The page supports search, filtering by status, refresh, and switching between card and list views. It also provides entries for onboarding Agents and creating Agents. For details about creating an Agent, see "Create an Agent" in this topic.
-
Each Agent card shows the Agent name, status, creation method, creation time, Agent ID, specifications, Harness type and version, model, update pending hint, and description. If no description is provided, "No description" is displayed.
-
The bottom of the list shows the total number and pagination, and you can select the number of items per page.
-
-
Click the Agent name to go to the details page of the Agent.
Common information on the details page
-
The top of the page shows Back, the Agent name, status, a version switch drop-down list (workspace version by default), and the creation method tag. The upper-right corner provides Edit configuration andDelete.
-
The Agent group in the left-side navigation contains eight pages: Overview and configuration, Debugging, System prompt, Endpoints and versions, Instances and Sessions (the badge shows the number of instances), API access, External channel configuration, and Observability.
Overview and configuration
The Overview and configuration page displays the following information in read-only mode and provides two operation entries: Edit context and Manage Agent assets.
|
Section |
Displayed content |
|
Basic information |
You can refresh in the upper-right corner. Displays the Agent ID, deployment method, creation method, creation time, model connection, model, reasoning effort, configuration version, Harness, Harness version, and team. When instances need to be updated, an "N instances to update" entry is displayed. |
|
Resource configuration |
Specifications, number of parallel Sessions, Session TTL (s), automatic scaling, minimum number of nodes, and maximum number of nodes. |
|
Network configuration |
Allow the default network interface to access the internet, Allow VPC access, and the enabled or disabled status of each. |
|
Agent context |
Provides three tabs: Memory, Knowledge base, and Data connection. Each tab shows the number of bound items. When none are bound, an empty state is displayed, for example, No memory bound yet. Click Edit context to bind items. |
|
Agent assets |
Provides four tabs: Skill, Tools, SubAgent, and Credential. Each tab shows the number of bound items. When none are bound, an empty state is displayed, for example, No skills bound yet. Click Manage Agent assets to manage them. |
Debugging
-
On the Debugging page, select an endpoint from the Select endpoint drop-down list (
defaultis selected by default), select a protocol from the protocol drop-down list (for example, AGUI), and then click Create session. -
In the status area, view the status and name of the debugging endpoint and the session ID that can be copied.
-
Enter a question or instruction in the input box. The placeholder is "Describe your question or instruction". Press Enter to send and Shift+Enter for a new line. The system starts a conversation with the Agent using the selected protocol and supports multi-turn context, tool calls, and chain-of-thought display.
The bottom of the page displays: "AI may generate incorrect information. Use your judgment."
System prompt
The System prompt page is used to view and configure the system prompt of the Agent. If no system prompt is configured, "No system prompt configured" is displayed.
Endpoints and versions
On the Endpoints and versions page, the Endpoint tab is titled Endpoint information and supports search by name or address, refresh, and endpoint creation.
The endpoint list contains the following columns:
|
Column |
Description |
|
Name/Status |
Endpoint name and status. |
|
Bound version |
The version bound to the endpoint. |
|
Access address |
Displayed in two rows: public network and VPC. The public network address can be copied. If VPC is not configured, |
|
Instance capacity |
The instance capacity of the endpoint. |
|
Actions |
Details, Edit, and Elastic configuration. |
The default endpoint is named default, has the Available status and a Default tag, is bound to the workspace version, and its instance capacity inherits the Agent configuration.
Click Create endpoint and configure the following parameters in the dialog box:
|
Parameter |
Description |
|
Name |
Required. The name can contain 1 to 16 characters, including only lowercase letters, digits, hyphens ( |
|
Bound version |
Required. Select a version from the drop-down list. |
|
Public network access |
Switch. Enabled by default. |
|
Minimum number of instances |
Required. Default value: 1. Valid values: 0 to 100. |
|
Maximum number of instances |
Required. Default value: 1. Valid values: 0 to 100. |
|
Enable dynamic policy |
Switch. Disabled by default. |
This capacity applies only to the current endpoint. Other entries are configured separately.
Before you use this feature:
-
When a dynamic policy is configured and is within its validity period, the minimum and maximum number of instances configured above do not take effect.
-
While a high-priority policy is in effect, regular policies do not take effect. The validity periods of different high-priority policies cannot overlap.
-
When multiple regular policies take effect at the same time, the maximum value of the minimum number of instances and the maximum value of the maximum number of instances are used.
Instances and Sessions
-
On the Instances and Sessions page, select an endpoint from the Select endpoint drop-down list at the top. This is required.
defaultis selected by default, and the instance list is filtered by the selected endpoint. -
You can search by instance ID or Session ID and refresh the list. After you select instances, you can click Update to latest version to update them in batches. If no instance is selected, the button is unavailable.
The instance list contains the following columns:
|
Column |
Description |
|
Checkbox |
Used to select instances. |
|
Instance ID |
The ID of the instance. |
|
Status |
The instance status. |
|
Number of Sessions |
Current value/maximum value. The maximum value corresponds to the number of parallel Sessions. |
|
Running version |
The version that the instance is running. If an update is pending, a pending update tag is displayed. |
|
Actions |
View Sessions and Update. |
API access
On the API access page, you can invoke the Agent from external systems through the AG-UI endpoint and configure the inbound access security policy. For details, see the "API access" section in this topic.
External channel configuration
-
The top of the page displays: To connect an Agent to an IM channel, read the integration guide first. You can click View documentation to open it.
-
Supports channel search and refresh.
-
The page provides three channel cards: DingTalk, Feishu/Lark, and WeCom. Each card shows the Not configured or Configured status and explains that the Agent can be accessed through the bot of the corresponding channel.
-
For the detailed configuration process, see Connect an Agent to an IM channel.
Observability
The left-side navigation provides the Observability page.