---
title: "Agents"
description: "Learn about creating and configuring individual agents within your multi-agent system"
icon: "robot"
---

## Overview

Agents are the core building blocks of Rowboat's multi-agent system. Each agent carries out a specific part of a conversation, handles tasks via tools, and can collaborate with other agents to orchestrate complex workflows.

They are powered by LLMs and can:
- Respond to user input
- Trigger tools or APIs
- Pass control to other agents using @mentions
- Fetch or process internal data
- Execute RAG (Retrieval-Augmented Generation) queries
- Participate in sequential pipeline workflows

---

## Agent Types

Rowboat supports several types of agents, each designed for specific use cases:

| Name | Purpose | Characteristics |
|------|---------|-----------------|
| **Conversational Agents** (`conversation`) | Primary user-facing agents that interact directly with users and orchestrate workflows. | • Can respond to users and orchestrate workflows<br />• Typically serve as the start agent (Hub Agent)|
| **Task Agents** (`internal`) | Specialized agents that perform specific tasks without direct user interaction. | • Focused on specific functions<br />• Return results to parent agents|
| **Pipeline Agents** (`pipeline`) | Sequential workflow execution agents that process data in a chain. | • Execute in sequence within a pipeline<br />• Cannot transfer to other agents directly|


---

## Agent Configuration

Agents are configured through two main tabs in the Rowboat Studio interface:

### **Instructions Tab**


#### Description
A clear description of the agent's role and responsibilities

#### Instructions
Instructions are the backbone of the agent's behavior. Use the Copilot's structured format for consistency:

**Recommended Structure:**
```
## 🧑‍💼 Role:
[Clear description of the agent's role]

## ⚙️ Steps to Follow:
1. [Step 1]
2. [Step 2]
3. [Step 3]

## 🎯 Scope:
✅ In Scope:
- [What the agent should handle]

❌ Out of Scope:
- [What the agent should NOT handle]

## 📋 Guidelines:
✔️ Dos:
- [Positive behaviors]

🚫 Don'ts:
- [Negative behaviors]
```

#### Examples
These help agents behave correctly in specific situations. Each example can include:
- A sample user message
- The expected agent response
- Any tool calls (if applicable)

### **Configurations Tab**

#### Name 
Name of the agent


#### Behaviour
- **Agent Type**: Choose from `conversation`, `internal`, or `pipeline`
- **Model**: Select the LLM model (GPT-4.1, GPT-4o, google/gemini-2.5-flash, etc.)

#### RAG
- **Add Source**: Connect data sources to enable RAG capabilities for the agent

---

## Creating Your Initial Set of Agents

Let Copilot bootstrap your agent graph.

### Instruct Copilot

Start by telling Copilot what your assistant is meant to do — it'll generate an initial set of agents with best-practice instructions, role definitions, and connected agents.

<Frame>
  <img src="/docs/img/create-agents-delivery.png" className="w-full max-w-[400px] rounded-xl" alt="Creating agents with Copilot" />
</Frame>

### Inspect the Output

After applying the suggested agents, take a close look at each one's:
- **Instructions**: Define how the agent behaves
- **Examples**: Guide agent responses and tool use

<Frame>
  <img src="/docs/img/agent-instruction.png" alt="Inspect agent instructions" />
</Frame>

---

## Updating Agent Behavior

There are three ways to update an agent:

### 1. With Copilot

Copilot understands the current chat context and can help rewrite or improve an agent's behavior based on how it performed.

<Frame>
  <img src="/docs/img/update-agent-copilot.png" className="w-full max-w-[400px] rounded-xl" alt="Update agent using Copilot" />
</Frame>



### 2. Manual Edits

You can always manually edit the agent's instructions.

<Frame>
  <img src="/docs/img/update-agent-manual.png" alt="Manually edit agent" />
</Frame>

---