Core Concepts

Introduction

Bumblebee is an AI Workflow Builder that lets you build powerful workflows in natural language. To get the most out of Bumblebee, it's important to understand its core building blocks.

Core Entities

Node

A node is a single step in a workflow. It represents the smallest unit of operation.

Edge

An edge is a connection between two nodes. It determines the next step after a node finishes running.

Workflow

A workflow is a reusable process that achieves a business objective consistently and automatically. It defines a process and typically maps to the smallest reusable procedure.

Practically, a workflow is defined as a list of nodes and edges connecting between them. A workflow must be a Directed Acyclic Graph (DAG), meaning it cannot contain loops between nodes.

WorkflowRun

A workflow run is a single execution of a given workflow definition. When a workflow run starts, users can provide free-form input, which may lead the workflow run down different paths.

When a workflow run is triggered, it follows the workflow's definition at that point in time. If a workflow definition changes while a run is ongoing, the run will continue to follow the original workflow definition until completion.

Workflow runs are stored and available for 90 days, after which they are automatically cleaned up.

Types of Nodes

There are different types of nodes. Each node type is designed for a fundamentally different purpose.

Agent Node

An agent node is a Large Language Model (LLM) powered AI agent that achieves a given objective using tools. It is the most recommended node type to use by default.

Users control the objective of an agent node through a natural language instruction. The instruction typically highlights the desired outcome and may describe how to achieve it. Instructions can also include guardrails and edge case handling.

Users often need to give an agent node access to tools to unlock its full power. Tools are pre-built integrations that users can configure to interact with their own data. Under the hood, a tool is one or multiple API calls maintained by Bumblebee. See Integrations for a full list of available integrations.

Agent nodes have a series of default configurations, such as LLM provider and model name. We don't currently support bringing your own LLM, but may open up this capability in the near future.

Condition Node

A condition node is a Large Language Model (LLM) powered boolean (true or false) evaluator. Users supply a natural language condition, which the node uses to evaluate against its input.

Instead of having one outgoing edge, a condition node has two outgoing edges: one for when the condition evaluates to true and one for when it evaluates to false. This node enables workflow branching to support more complex processes.

How Data Flows Through a Workflow

When a workflow run starts, users supply an input object as the workflow run input. The input is fed into the first node (the starting node of the Directed Acyclic Graph). After execution, the first node produces an output, which is then fed as input to the next node. This sequence continues until the last node finishes running. The output of the last node becomes the output of the entire workflow run.