This topic introduces the core concepts of AgentCore, including the relationships among Workspace, Agent, model, Model Connection, Team, Skill, and MCP.
Resources in AgentCore are organized with the Workspace as the boundary. Within a Workspace, model configurations, Agents, Teams, and MCP tools are maintained independently. An Agent is the unit that executes AI tasks; at runtime it is associated with a model and can be bound with Skills and MCP tools. Multiple Agents can form a Team.
Concept overview
Concept | Definition |
AgentCore | The AI agent building and management platform of Alibaba Cloud. |
Workspace | The basic unit of resource isolation and management. Models, Agents, Teams, and MCP tools are maintained independently without mutual interference. |
Agent | The execution unit of an AI agent. Created within a Workspace, it receives tasks, invokes a model to complete reasoning, and returns results. Its capabilities are determined by the bound assets: binding Skills provides professional skills, binding MCP tools provides external tool-calling capabilities, and models and credentials provide reasoning capability and the keys required to access external services. |
Agent template | A reusable Agent configuration template for quickly creating the same type of Agent, reducing repeated configuration. |
Agent asset: Skill | The professional skill invoked by an Agent. Three import methods are supported: ZIP upload, Nacos CLI, and preset templates. |
Agent asset: MCP | Model Context Protocol. In AgentCore, it extends standardized tool integration for Agents. Two types, HTTP to MCP and direct proxy, both require network reachability. |
Model Connection | The channel for connecting to model providers. On the connection side, the model provider, API endpoint, protocol, and access credentials are configured. |
Model | A reasoning model selectable by Agents under a connection. On the model side, the context window, maximum output tokens per request, and streaming output are recorded, along with capability items such as tool use, vision, and audio. |
Credential | A key required to access external services, centrally stored as key-value pairs. The Key is the environment variable name (such as GITHUB_ACCESS_TOKEN), and the Value is hidden by default after creation. Credentials are referenced by the authentication configurations of Model Connections and tools. |
Concept details
AgentCore
AgentCore is the AI agent building and management platform of Alibaba Cloud. Its capabilities are divided into four groups: Agent building, Asset Management, Agent collaboration, and identity and permissions. Resources in AgentCore are organized with the Workspace as the boundary. Each Workspace independently maintains model configurations, Agents, Teams, and MCP tools, ensuring isolation between different business scenarios.
Workspace
A Workspace is the basic unit of resource isolation and management in AgentCore, supporting independent resource allocation by project or business line.
For more information, see Manage Workspace.
Agent
An Agent is the execution unit of an AI agent in AgentCore. It is associated with a model and executes specific AI tasks.
Agent type | Description |
Harness Agent | An Agent whose runtime framework is provided by AgentCore. It can be created and run after configuration in the console, without self-development or deployment. Configuration items include the Agent name, model, Team affiliation, SOUL.MD and AGENT.MD files, Skill bindings, and MCP tool bindings. |
High-code Agent (coming soon) | An Agent developed by writing code, suitable for scenarios that require custom business logic and fine-grained control over the running process. |
Managed local Agent | Integrates an Agent deployed on the customer's premises into AgentCore for unified management. It can participate in team orchestration and collaboration without modification. |
Managed Agent (coming soon) | Registers an Agent deployed outside AgentCore to the platform for unified scheduling. |
The following items are configured when creating a Harness Agent:
Configuration item | Description |
Agent name | The name of the Agent, used to identify the Agent in the Agent list. |
Model selection | Specifies the model invoked when the Agent executes tasks, selected from the models configured in the Workspace. |
System prompt | Defines the role setting and behavioral guidelines of the Agent. The content is written into SOUL.MD and AGENT.MD respectively. |
SubAgent | Adds SubAgents to the Agent, splitting complex tasks among sub-Agents for separate processing, with the current Agent responsible for scheduling. |
Resource configuration | Sets the resource specifications used at Agent runtime. |
Skill binding | Binds Skills to supplement the Agent with domain-specific professional capabilities. |
MCP binding | Binds MCP tools to extend the Agent's capability to invoke external services and tools. |
For more information, see Manage Agent.
Agent template
An Agent template is a reusable Agent configuration for quickly creating Agents with the same configuration, reducing item-by-item repeated configuration.
For more information, see Manage Agent Templates.
Skill
A Skill is a professional skill invoked by an Agent, such as sending notifications, querying data, or triggering workflows. The content of a Skill is defined in SKILL.md, whose name and description fields correspond to the Skill name and description in the console respectively. Skills support three import methods: uploading a locally packaged Skill in ZIP format, importing a local Skill folder with Nacos CLI, and importing a preset Skill template.
For more information, see Manage Skill.
MCP
MCP (Model Context Protocol) is a protocol for connecting Agents with external tools and data sources. In AgentCore, MCP extends standardized tool integration capabilities for Agents, used to connect external tools or third-party data sources.
For more information, see Manage MCP Servers.
Team
A Team is an organizational form composed of multiple Agents, adding one more layer of organization on top of Agents.
For more information, see Set up Team Collaboration.