AI Users Clarify Agent Terminology
Core Concepts
Message: Text instructions (input from user or system to Agent).
Agent: An intelligent entity based on large models (LLM), equipped with tools, connectors, and sometimes running on a computer, capable of receiving messages, executing tasks, and returning results.
Local Agent: A computer running on the user's local desktop (high privacy, low latency).
Cloud Agent: A computer running on the internet cloud (powerful computing, scalable).
Subagent: The main Agent delegates part of the tasks to another Agent (multi-agent collaboration).
Sandbox: A real or virtual isolated environment where the Agent operates, ensuring safety for both the user and the Agent.
Capabilities and Extensions
Tool: Specific input/output functions provided by AI (e.g., search, code execution, image generation).
Skills: A reusable set of instructions + code that enables the Agent to master fixed capabilities.
Plugin: A packaged combination of Skills + Connectors.
Harness: A highly structured collection of instructions, tools, and processes that helps the Agent maximize the use of models and computers to complete complex tasks (especially coding).
Agent SDK: A development toolkit for programmatically calling Harness.
Connections and Triggers
Connector: A bridge for bidirectional information exchange between the Agent and third-party systems.
MCP: A standardized protocol for Connectors (facilitating ecosystem interoperability).
API: A standard way for software to communicate through code.
CLI: Triggering code (including API) through terminal commands, more reliable than simulating button clicks by the Agent.
Hooks: Specific event triggers in the Agent's lifecycle (e.g., receiving a certain type of message, task completion, etc., for automatic execution).
Overall Architecture Understanding
A mature Agent system typically consists of:
Agent (LLM + Harness)
↓ Uses
Tools / Skills / Plugins + Connectors (MCP standard)
↓ Interacts with the external world through
API / CLI
↓ Runs in
Sandbox (Local / Cloud)
↓ Supports
Subagent collaboration + Hooks event-driven