AgentChat

AgentChat empowers developers to create AI assistants, providing the capability to build sophisticated multi-agent systems. Its features include:

• Boost Autonomy

  AgentChat differ from those in systems like AutoGen. With built-in tools, they operate independently, eliminating the need to rely on other agents for information. This allows them to iterate on a task before delivering results or involving another agent, reducing unnecessary interactions and agents.

• Tool Integration

  Unlike LangChain or AutoGen, which require following serval rules to register functions or tools, AgentChat allows you to simply write a function and add a comment. The agent will automatically use the tool to solve tasks. As shown in the example below, check out sample/agent_tool.py for a quick start.

  Agent(
    model_client,
    "Assistant AI", "You are an assistant to solve tasks",
    tools=[wikipedia],
  ).run()
  

• Governed Actions

  Actions performed by the AgentChat are regulated by developers with three permission levels:

  • auto: Requests permission only for actions that modify the system or environment[PromptAgent]
  • always: Requests permission for every action.
  • none: Never requests permission.

• Multi-Agent System

  Transitioning the chat-agent into a multi-agent system a straightforward process. The handoff workflow for orchestrating agents draws inspiration from the post Routines and Handoffs, which details the functionality of the Swarm project. We strive to achieve a harmonious balance, enabling you to create a single agent for specific tasks while effortlessly evolving towards a sophisticated multi-agent framework.

• Memory [PROCESSING]

  Memory capabilities enhance accuracy and optimize thought processes by transitioning from stateless to stateful operations. Unlike Retrieval-Augmented Generation (RAG), which builds knowledge from external sources, our approach is based on the agent's own experiences.

  Zen-Agent provides an interface called ChatMemory that allows you to customize memory for your assistant. We offer two default memory implementations:

  1. ChatBufferMemory A short-term memory solution designed to retrieve the most recent message along with the current session context.

  2. ChatVectorMemory A long-term memory implementation based on LlamaIndex vector memory.

  > MemGPT: Towards LLMs as Operating Systems > CLIN: A CONTINUALLY LEARNING LANGUAGE AGENT FOR RAPID TASK ADAPTATION AND GENERALIZATION

  3. ChatPgMemory ...

• RAG Support

We also provide a retrieval agent capable of integrating local resources or knowledge into the multi-agent system. The default implementation is based on LlamaIndex's ChatEngine.

• Typed Prompt and Auto Optimizer

  • https://github.com/stanfordnlp/dspy
  • https://github.com/zou-group/textgrad

• UI Support

  • https://github.com/streamlit/streamlit