What an agent is made of

An agent is one AI specialist with a role. This page explains what an agent is made of, so that the other pages of this section are easier to follow.

In these pages, the example is an agent named Reminder Writer. It writes a polite payment reminder for a customer who has not paid an invoice.

The three things every agent needs

Part	What it is	Where you set it
Role	One or two sentences on what the agent is good at. A team reads it to decide when to use this agent.	The panel Info of the editor.
Instructions	What the agent must do and how, in plain words.	The window of the AI Agent node, field System instructions.
Model	The AI model the agent thinks with.	A model node, linked to the agent.

See Create an agent.

What an agent looks like in the Studio

In its editor, an agent is drawn as one large box, AI Agent, with a row of seven diamonds under it. Each diamond is a place where you plug something into the agent. What you plug in is drawn as a circle, linked to the diamond by a dotted line.

The seven diamonds, from left to right

#	Name	What you plug in	Needed?
1	LLM Model	The AI model of the agent. One only.	Yes
2	Knowledge Bases	The documents the agent relies on, each with a Knowledge Base step. Several are possible. See Give an agent knowledge and a memory.	No
3	Human Loop Channel	The way the agent reaches a person when it needs an answer. One only.	No
4	Agent Session	A place to keep the conversation, so the agent remembers what was said before. One only.	No
5	AI Agent Tools	Tools the agent can use: a calculator, a web search, a data table. Several are possible.	No
6	AI Agent Policies	Rules applied to the way the agent runs.	No
7	AI Agent Guardrails	Checks on what goes in and out of the agent, for example to block a word.	No

To choose between instructions, a Data Table, a Knowledge Base, a session and a Vector Store, see Where each kind of information goes.

An agent with a model and good instructions is already useful. Add the rest only when the job needs it.

How an agent works: the execution mode

The window of the AI Agent node has a setting, Agent execution mode. It has three values:

Mode	How the agent works
Tools	It answers, and it may call the tools you plugged in, choosing which one to use. This is the default.
Conversational	It answers directly from the conversation, without using any tool.
ReAct	It thinks, uses a tool, looks at the result, and repeats until it can answer.

Where an agent is used

Alone. You open it and give it a task. See Run an agent and read its answer.
In a team. It becomes a member, next to other agents. See Create a team of agents.
In a workflow. It becomes one step among fixed steps. See Add an AI step to a workflow.

Next step

Create an agent.