Sub-Components¶

Sub-components are configuration and utility nodes that attach to AI nodes via special diamond-shaped handles. They do not appear as standalone steps in the execution flow -- instead, they provide capabilities that AI nodes consume at build time or during their reasoning loop.

Overview¶

There are three sub-component types:

Component	Purpose	Connects Via
AI Model	Provides LLM credential and model selection	Blue diamond handle (`llm`)
Output Parser	Parses raw LLM output into structured data	Slate diamond handle (`output_parser`)
Code Execute	Executes Python or Bash in a sandboxed subprocess	Green diamond handle (`tool`)

How sub-components work¶

Sub-components connect to AI nodes via the colored diamond handles at the bottom of the node. The connection type determines how the sub-component integrates:

Handle	Color	Edge Label	Description
model	Blue (`#3b82f6`)	`llm`	AI Model provides the LLM configuration
output_parser	Slate (`#94a3b8`)	`output_parser`	Output Parser processes raw LLM text
tools	Green (`#10b981`)	`tool`	Code Execute provides a tool function

AI Model¶

The AI Model node is unique among sub-components:

It has only a top diamond handle (source) -- it connects upward to nodes that need a model
It is non-executable -- it does not run during execution
It is required for all AI nodes (agent, categorizer, router, extractor)

Without an AI Model connected, an AI node cannot resolve which LLM to use and will fail at build time.

Output Parser¶

The Output Parser takes the raw text output from an AI node and transforms it into structured data (JSON, regex matches, or lists). It is:

Non-executable -- it runs as part of the AI node's post-processing, not as a separate execution step
Used by categorizer, router, and extractor nodes (not agents)

Code Execute¶

The Code Execute node provides a sandboxed code execution tool to agents. Unlike the other two sub-components, it is:

Executable -- it runs when the agent's LLM invokes it during the reasoning loop
Connected via the tools handle (green diamond), same as other tool nodes
A LangChain @tool that accepts code and language parameters

Canvas appearance¶

Sub-components are visually smaller than regular nodes. AI Model nodes sit below AI nodes and connect upward. Output Parser and Code Execute nodes connect to the bottom handles of their parent AI node.

flowchart TB
    A[Agent Node]
    M[AI Model] -.->|model| A
    OP[Output Parser] -.->|output_parser| A
    CE[Code Execute] -.->|tool| A

Non-executable nodes

AI Model and Output Parser are marked as non-executable in the node type registry. They do not show running/success/failed status badges during execution. Code Execute, as a tool, does show execution status when invoked by an agent.