There are a few key concepts in AutoChain, which could be easily extended to build new agents.
Chain is the overall stateful orchestrator for agent interaction. It determines when to use
tools or respond to users. Chain is the only stateful component, so all the interactions with
memory happen at the Chain level. By Default, it saves all the chat conversation history and
intermediate AgentAction with corresponding outputs at prep_input and prep_output steps.
Agent provides ways of interaction, while Chain determines how to
interact with agent.
Read more about the chain concept.
Agent is the stateless component that decides how to respond to the user or whether an agent
requires to use tools.
It could contain different prompts for different functionalities an agent could have. The main goal
for an agent is to plan for the next step, either respond to the user with AgentFinish or take an
action with AgentAction.
Read more about agent.
The ability to use tools make the agent incredible more powerful as shown in LangChain and AutoGPT. We follow a similar concept of "tool" as in LangChain here as well. All the tools in LangChain can be easily ported over to AutoChain if you like, since they follow a very similar interface.
Read more about tool.
It is important for a chain to keep the memory for a particular conversation with a user. The
memory
interface exposes two ways to save memories. One is save_conversation which saves the chat
history between the agent and the user, and save_memory to save any additional information
for any specific business logics.
By default, memory are saved/updated in the beginning and updated in the end at Chain level.
Memory saves conversation history, including the latest user query, and intermediate
steps, which is a list of AgentAction taken with corresponding outputs.
All memorized contents are usually provided to Agent for planning the next step.
Read more about memory