Berkeley LLM Agent Lecture 03: Intro to Agentic AI and AutoGen & Future of Knowledge Assistants
Series Paper Reading
Paper 1: AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
Define agents and then get them to talk.
AutoGen is an open-source framework that allows developers to build LLM applications via multiple agents that can converse with each other to accomplish tasks.
AutoGen has two key concepts of AutoGen: conversable agents and conversation programming. As for conversation programming, two core concepts are raised here: computation and control flow.
For computation, it means the actions agents take to compute their response in a multi-agent conversation.
For control flow, the sequence under which these computations happen, following two design patterns to facilitate conversation programming: Unified interfaces and auto-reply mechanisms for automated agent chat, control by fusion of programming and natural language.
Sample use cases are listed with AutoGen:
Autonomous math problem solving by directly reusing two built-in agents from AutoGen
Retrieval-Augmented Code Generation and Question Answering
Multi-Agent Coding
Dynamic Group Chat
Conversational Chess
Paper 2: StateFlow: Enhancing LLM Task-Solving through State-Driven Workflows
StateFlow is a novel LLM-based task-solving paradigm that conceptualizes complex task-solving processes as state machines. “Process ground” (via state and state transitions) and “Sub-task solving” (through actions within a state), enhancing control and interpretability of the task-solving procedure.