Agents

Define AI agents that combine models, prompts, and tools for specific tasks.

What Are Agents?

An agent is an AI assistant configured to perform a specific task. Each agent brings together three things: a model (which LLM to use), a prompt (the instructions that define behavior), and a set of tools (the capabilities the agent can use). You create and configure agents entirely through the admin UI — no code deployment needed.

Think of an agent as a specialist you hire for a particular job. A "Data Analyst" agent might use a powerful reasoning model, have instructions for interpreting business metrics, and be equipped with tools to query your data warehouse and generate charts. A "Content Editor" agent might use a fast, cost-effective model with instructions focused on grammar and style, equipped with tools to search your content library.

Agent Configuration

Every agent is defined by a small set of properties that you configure in the admin UI:

Property	Purpose
Name	A human-readable name that identifies the agent in the UI and in chat interfaces. Choose something descriptive — "Customer Support Agent" is better than "Agent 1."
Key	A stable identifier used to reference the agent programmatically. Keys remain consistent across environments (development, staging, production).
Description	A brief explanation of what the agent does and when to use it. This helps team members understand the agent's purpose at a glance.
Model	Which LLM the agent uses for generating responses. You select from the models configured in your system. See Models for details.
Prompt	The instructions that define the agent's behavior — its persona, rules, and response format. See Prompts for details.
Tools	The capabilities assigned to the agent. Tools let the agent take actions beyond generating text — querying data, searching content, generating images, executing scripts, and more.

Tool Association

Tools are what make agents useful beyond simple conversation. By assigning tools to an agent, you give it the ability to take real actions during a chat session. The agent decides when to use a tool based on the user's request and the instructions in its prompt.

Available tool types include:

SQL queries — Run queries against your analytics data warehouse to retrieve metrics, trends, and reports.
Content search — Search the platform's content library by text, metadata, categories, or traits.
Image generation — Generate images from text descriptions using AI image models.
Script execution — Run predefined scripts to perform custom operations.
MCP servers — Connect to external tools and data sources via the Model Context Protocol. See MCP Servers for details.

You can assign multiple tools to a single agent. The agent will choose the right tool for each situation based on its prompt instructions and the user's request.

Start with a small set of tools and expand as needed. An agent with too many tools may struggle to choose the right one. If an agent needs many capabilities, consider breaking it into sub-agents with focused tool sets.

Hierarchical Composition

Agents can have sub-agents, enabling complex multi-step workflows. A parent agent can delegate parts of a task to specialized child agents, each with their own model, prompt, and tools. This is the key to building sophisticated AI workflows from simple, manageable components.

Consider a "Research Agent" workflow:

The user asks the Research Agent to investigate a topic.
The Research Agent delegates to a Search Agent to find relevant content and data sources.
The Research Agent passes those results to a Data Analysis Agent to run queries and identify patterns.
Finally, the Research Agent sends everything to a Summarization Agent to produce a clear, concise report.

Each sub-agent is independently configurable. You can swap the model on the Summarization Agent without affecting the others. You can reuse the Search Agent as a sub-agent of a completely different parent agent. This composability makes the system flexible and maintainable as your needs grow.

Sub-agents inherit no configuration from their parent. Each agent — whether top-level or nested — has its own model, prompt, and tool assignments. This keeps behavior predictable and makes each agent independently testable.

Chat Interface

Users interact with agents through a real-time streaming chat interface. Messages are delivered token by token as the model generates them, so users see responses forming in real time rather than waiting for a complete answer.

Each conversation with an agent creates a chat session. Sessions are persistent — they maintain the full message history so users can return to a conversation later, review previous exchanges, or continue where they left off. Every session tracks:

Title — A name for the conversation, which can be auto-generated based on the discussion topic.
State — Whether the session is active, completed, or archived.
Message history — The complete record of user messages, agent responses, and tool interactions.

Interactive Elements in Chat

Agents are not limited to plain text responses. During a conversation, agents can present interactive elements that make the experience richer and more useful:

Element	How It Works
Confirmation requests	The agent pauses and asks the user to approve an action before executing it. This is useful for potentially impactful operations like running a database query that modifies data or publishing content.
Choice selection	The agent presents a set of options for the user to choose from. This guides the conversation and ensures the agent has clear direction for the next step.
Charts and visualizations	The agent can render data as bar charts, line charts, pie charts, and other visual formats directly in the chat. This is especially useful for data analysis agents presenting query results.
Tables	Structured data can be displayed in formatted tables within the chat, making it easy to scan and compare information.

Built-In Content Agents

The platform includes several pre-configured agents that work directly with the content management system. These agents automate common content operations and can be used as-is or customized to fit your needs:

Auto-generate descriptions — Reads the body of a content item and produces a concise summary suitable for metadata, search results, or social sharing.
Extract topics — Analyzes content to identify and tag the key subjects, themes, and concepts covered.
Estimate reading time — Calculates how long it will take an average reader to consume a piece of content, factoring in text length and complexity.
Generate content from specs — Takes a specification, outline, or brief and produces a full content draft following the structure and guidelines provided.

These built-in agents are a great starting point for learning how agents are configured. Open one in the admin UI to see how its model, prompt, and tools are set up, then use that as a template for your own agents.

Best Practices for Agent Design

One agent, one job. Agents work best when they have a clear, focused purpose. Instead of a "do everything" agent, create specialized agents and compose them with the hierarchical system.
Match the model to the task. Use fast, inexpensive models for simple tasks like classification and extraction. Reserve powerful reasoning models for complex analysis and generation.
Write clear prompts. The prompt is the most important factor in agent quality. Be specific about what the agent should and should not do. See Prompts for guidance.
Test with real conversations. After configuring an agent, have a real conversation with it. Try edge cases, unexpected inputs, and multi-turn exchanges to see how it behaves.
Iterate without fear. Since everything is configured in the UI, you can change a prompt, swap a model, or add a tool and immediately see the effect. There is no deployment cycle to slow you down.