Agent Steps

Conduet agents are built by composing steps on a visual canvas. Each step handles one concrete task — talking to the user, branching on a condition, calling an API, or delegating to a sub-agent — and you connect them together to shape the conversation.

Core steps

These are the building blocks most agents use. Start with an Agent step for open-ended reasoning, add Messages and Buttons for structured moments, and reach for KB Search and API Call when you need outside information.

Agent

A multi-turn conversational agent with its own instructions, model, and tool access. The agent decides how to respond, when to use tools, when to search the knowledge base, and when to hand off — just like a human specialist taking a conversation end-to-end.

Message

Send a scripted or generated message to the user. Use scripted messages for fixed copy like greetings and disclaimers, or generate responses dynamically when the tone should match the conversation.

Buttons

Present the user with a set of quick options. Each button can route the conversation down a different path or send a text value — perfect for menus, confirmations, and guided choices.

Condition

Branch the conversation based on variables collected so far. Combine multiple conditions with AND/OR logic, and fall back to an else path when nothing matches.

Knowledge Base Search

Find the most relevant passages in your knowledge base and inject them as context for the next response. Great for grounding answers in your own documentation without having to write them into a prompt.

API Call

Make an HTTP request to any external service — CRM, booking system, internal tool — and map the response into variables the agent can use. Supports GET, POST, PUT, PATCH, and DELETE with full header and body control.

Playbook

Hand off a portion of the conversation to a reusable sub-agent. Playbooks run until they collect what they need, then return control to the parent workflow with the results.

Beyond the basics

More advanced step types are available for orchestrating multiple agents, running parallel branches, capturing validated input, displaying rich cards and carousels, executing custom code, and handing conversations off between workflows. Drop any of them onto the canvas from the step palette when you need them — they are documented in the builder itself as you configure them.