Key Concepts

A project is the top-level container for everything related to a single AI agent. It holds workflows, playbooks, tools, variables, secrets, knowledge base documents, widget configuration, and deployment settings. Each project has a unique slug and can be published as versioned snapshots.

Every project uses one of three frameworks. "Agentic" creates autonomous agents that reason with LLMs, use tools, and decide their own next actions. "Conversation Flow" builds structured, step-by-step dialogs where you define every path explicitly on the canvas. "Outbound" creates scheduled content pipelines that generate, approve, and publish posts to social channels (LinkedIn, Twitter, email, Slack, Discord, or webhooks). Choose agentic for open-ended tasks, conversation-flow for predictable scripted interactions, and outbound for automated content distribution.

A workflow is a directed graph of steps connected by edges. Each project can have multiple workflows. Steps are placed on a visual canvas with x/y positions and connected via ports. The runtime engine traverses these connections, executing each step in sequence or following conditional branches.

Steps are the individual units of logic within a workflow. Each step has a type (agent, message, prompt, condition, API call, etc.), configuration data, input/output ports, and a position on the canvas. Steps are connected by edges that define the conversation flow.

Connections (edges) link one step's output port to another step's input port. They define the execution order and control flow. Conditional steps and button steps can have multiple output ports, enabling branching paths through the workflow.

Playbooks are reusable agent configurations with their own instructions, model settings, tools, and exit conditions. They act like sub-agents that can be invoked from workflows. When a playbook's exit condition is met (e.g., required variables are collected), control returns to the calling workflow.

Tools extend what your agent can do. There are four types: API tools make HTTP requests to external services, function tools execute custom code, integration tools connect to third-party platforms, and MCP tools communicate with Model Context Protocol servers. Agents can call tools during conversations to fetch data, perform actions, or access external systems.

Variables store dynamic data during conversations — user input, API responses, computed values. They have names and optional default values, and are accessible across all steps in a workflow. Secrets are encrypted key-value pairs for sensitive data like API tokens. They are never exposed in responses or logs.

The knowledge base is a collection of documents (text, PDF, DOCX, URLs) that your agent can search during conversations. Upload content once and Conduet indexes it for semantic retrieval — during a conversation the agent automatically surfaces the most relevant passages and uses them to ground its answers.

A session represents one user's ongoing conversation with your agent. It stores the current state including variable values, conversation history, the agent's position in the workflow, and any internal execution state. Sessions are identified by a user ID and persist across multiple interactions.

When you publish a project, a version snapshot is created that captures the complete project state at that point in time. Deployments reference a specific version. You can target "development" (latest unpublished changes), "production" (most recent published version), or a specific version ID.

A deployment connects your published agent to a communication channel. Supported channels include webchat (embeddable widget), WhatsApp, Facebook Messenger, Instagram, and WeChat. Each deployment has its own credentials, branding configuration, and is tied to a specific project version.