▾
This glossary defines concepts used when building agentic mapping applications with AI components in the ArcGIS Maps SDK for JavaScript. It covers how user requests are interpreted and routed, how agents carry out tasks, and how models, tools, and information support that work.
Agent
An executable unit that handles a request or task using defined behavior and available capabilities. It may be deterministic, model-driven, or a combination, and may run independently or as a step in a larger agent’s workflow.
Agent loop
An iterative process in which the agent chooses an action, uses a skill or tool, considers the result, and continues until it can complete the task.
Context
Information available while handling a request, rather than fixed authoring instructions. It can include application information, conversation history, and information explicitly shared by other agents. For example, the selected layer and current map extent can help interpret “summarize the features in this area.” The relevant information must be supplied to the agent.
Conversation history
Previous user and assistant messages in a conversation. This history can help interpret follow-up requests, such as “zoom to those features.” It is distinct from shared state, which contains data explicitly contributed by agents.
Graph agent
An agent whose execution is organized as a graph of work steps and transitions. The developer defines the structure in advance; runtime conditions, model decisions, or tool results can affect which of the defined paths runs.
Human-in-the-loop (HIL)
Pausing execution to request input from a user before continuing. The input can be approval or rejection, a selection from available options, or additional information needed to carry out a task.
Instructions
Guidance about what the agent should do, how it should behave, and constraints it should follow. Instructions can apply to the whole agent or to a particular skill.
Intent
What the user wants to accomplish with a request. The orchestrator and agents use an LLM to interpret a user’s intent. The LLM determines which agent should handle the request and what task it should carry out, including relevant tool calls and parameters. Authors of custom agents can include human-in-the-loop interrupts to give the user more control in explicitly conveying their intent.
Middleware
Middleware is code that runs at defined points to observe or change execution. Agent middleware can run before or after an agent executes and around model or tool calls. Tool middleware runs during a tool call and can inspect or change its inputs or result, or determine whether the call proceeds.
For example, human-in-the-loop middleware can pause execution to ask the user for approval, a selection, or additional information before continuing.
Node
A unit of work in an execution graph.
Orchestrator
Coordinates the handling of user requests across registered agents. It manages agent routing, carries conversation information forward, and handles execution results and interruptions. The orchestrator coordinates work across agents, while each agent defines how to carry out its assigned task.
Prompt
Input to a particular language-model call, potentially combining instructions, the user request, and relevant context. An agent’s configured prompt or template helps build that input and may be reused across calls. It does not automatically apply to other agents or steps.
Registration
The integration step that makes an agent available to the assistant. Registration includes the agent’s identity and a description of its capabilities, which helps the assistant select it for relevant requests. It is separate from how the agent’s internal behavior is designed.
Semantic search
Finding information by meaning rather than requiring exact text matches. Vector-based semantic search compares embeddings to identify relevant information, such as layers or fields that match a user’s request. The AI components uses semantic search to reduce the web map context it sends to an LLM to only the layers and fields most relevant to a user’s input.
Shared resource
An identifiable result an agent makes available for use by subsequent agents in a conversation, such as a point, polygon, or map extent. A shared resource has an ID, a description, and associated data, allowing later tasks to refer to and reuse it.
Shared state
Data explicitly shared by agents and carried forward between agent executions in a conversation. Subsequent agents can access it as part of their runtime context.
Skill
A markdown file outlining task-specific instructions, domain knowledge, and declarations of the tools it may use. A skill guides an agent but is not itself an agent.
Skills agent
An agent that uses an agent loop to choose its next action at runtime, guided by the request, available skills, and tool results. The sequence of skill and tool calls is not authored in advance, so the same task may follow different paths as context changes.
System prompt
Instructions that define a model’s role, behavior, and constraints. These instructions guide how it handles requests, while other parts of the prompt provide the current request and relevant information.
Tool
A function with additional metadata, such as a description, that may be invoked by an agent. Tool implementations can be used by one or more agents.
Vector embedding
A numerical representation of information that allows it to be compared by semantic similarity. In agentic mapping applications, embeddings can represent user requests and descriptions of map layers and fields to help find relevant information.
Workflow
An arrangement of steps and decisions for completing a task. In the AI components graph agent utilities, workflow classes compose agents into an executable unit that can be wrapped in a WorkflowAgent.