Overview

Quite often an agentic workflow needs further input from a human user (or another agent) in a muli-turn style conversation or verification.

Revisiting the figure from the background on agents, we can take a new view:

graph LR
    User[User] <-- "Input/Output" --> LLM[LLM Agent]
    LLM <--> Tools[Tools]
    LLM <-- "Iterative Input/Output" --> Environment[User, Other Users, Other Agents]
    User <-- "Iterative Input/Output" --> Environment

    style LLM fill:#FECACA, fill-opacity:0.3
    style User fill:#60A5FA, fill-opacity:0.3
    style Tools fill:#FBBF24, fill-opacity:0.3
    style Environment fill:#34D399, fill-opacity:0.3

To allow this for the users of the framework, we have designed an extendible abstract class called HIL (Human In the Loop). Additionally we have implemented a local chat server for quick development and prototyping of such behaviours. Please refer to the following sections for futher information:

• Local ChatUI
• Custom HIL Behaviour