Skip to main content

Overview

Our JSON configuration files make it easy to prototype agents quickly without having to code your own logic. These agents can complete automated workflows, as well as lead branching voice conversations based on pre-defined topics.

How the Finite State Machine Works​

The JSON configuration files use a finite state machine that evaluates the current state in order to select the agent's next set of actions (referred to as topics).

The State Machine guides to take a state as an input to determine the next state. In AgentStation, these states are referred to as topics, and they denote a series of actions that the agent can take. Through the JSON configuration file, the LLM is given context to select the next topic and run through all actions associated with it.

Agents built using the State Machine can take actions and select topics independently, or take input from a user through voice conversation over Zoom.

File Structure​

The JSON configuration file contains several sections:

  • Agent Context: These parameters give the LLM information on the problem they are trying to solve, the personality they should take, and how they should answer questions or objections from a user in conversation.
  • Topics: Topics contain a discreet set of tasks the LLM will complete if the topic is selected, and parameters that provide instructions to the LLM on how to select the next topic.
    • Tasks: Every topic contains at least one task, which are the actions the agent should take during that topic. Agents can speak, asl questions, complete browser actions, and interact in online meetings.

Getting Started​

We strongly recommend reviewing the example Github repo and modifying an existing JSON configuration file to start.