Skip to main content

Agent & MultiAgent

LizAbout 2 minLLMAgentMultiAgent

Agent & MultiAgent

  • Agent
  • MultiAgent

1. Agent

Agent = LLM + Observation + Thinking + Action + Memory

  • LLM: Processes information, makes decisions, and executes actions
  • Observation: Perception of the environment. An agent can receive messages from other intelligent entities in the environment.
  • Thinking: Analyzes the results of observations and memory content, considers what action to take next. This decision-making ability is provided by the LLM.
  • Action: The result of observation and thinking, determining what specific tasks to carry out. This is similar to tools in LangChain.
  • Memory: Stores past experiences.

Difference Between Agent and Chain

  • Similarities:

    • Both complete a series of tasks step by step to ultimately achieve complex goals.

  • Differences:

    • Chain: A predefined sequence of tasks, i.e., static (what should be done in each step is determined in advance).
    • Agent: The tasks in each step are not predefined and can vary depending on the user's questions or scenarios (dynamic).

2. ReAct

Paper Name: ReAct : SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS

Paper: https://arxiv.org/pdf/2210.03629

Github: https://github.com/ysymyth/ReAct

3. AutoGPT

Github: https://github.com/Significant-Gravitas/AutoGPT

Paper Name: Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions

Paper: https://arxiv.org/pdf/2306.02224

An autonomous agent only needs to be told the requirements, and does not need guidance on what to do first or next. It can autonomously propose a plan, execute it, and complete the task.

4. MetaGPT

Github: https://github.com/geekan/MetaGPT

Paper Name: METAGPT: META PROGRAMMING FOR A MULTI-AGENT COLLABORATIVE FRAMEWORK

Paper: https://arxiv.org/pdf/2308.00352

5. HuggingGPT

Paper Name: HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face

Paper: https://arxiv.org/pdf/2303.17580

Github: https://github.com/microsoft/JARVIS

Four Stages

  • Task Planning
  • Model Selection (from Hugging Face)
  • Task Execution
  • Response Generation

6. generative_agents: Stanford Town

Paper Name: Generative Agents: Interactive Simulacra of Human Behavior

Paper: https://dl.acm.org/doi/pdf/10.1145/3586183.3606763

Github: https://github.com/joonspk-research/generative_agents

This is one of the most influential works in this field, providing an effective foundation for future multi-agent collaboration.

There are many potential application scenarios to explore in the realm of generative multi-agents.

6.1. Design of Each Agent's Properties

  • Basic Information
  • Past Experiences
  • Interests and Hobbies

6.2. Memory Design

  • Recency (Time proximity):

    • The closer an event is to the present, the more weight it carries.

  • Importance:

    • Some events are more important than others, e.g., "Waking up" < "Promotion".

  • Relevance

    • Cosine Similarity

score = Recency x α1\alpha_1 + Importance x α2\alpha_2 + Relevance x α3\alpha_3

Reflection Extraction: Summarizing the Memory Stream

  • Scheduled Task: Summarize periodically after a certain time.

6.3. Planning: Generating a Daily Plan for Each Agent

  • Daily routine
  • Add unique schedules based on the characteristics of the agent

6.4. React: Interaction Between Agents

  • Whether interaction occurs
  • Whether the original plan needs to be changed

7. ChatDev: Multi-Agent Software Development

Paper Name: Communicative Agents for Software Development

Paper: https://openreview.net/pdf?id=yW0AZ5wPji

Github: https://github.com/OpenBMB/ChatDev

  • Define Roles
  • Memory Design

8. ToolLLM

Paper Name: TOOLLLM: FACILITATING LARGE LANGUAGE MODELS TO MASTER 16000+ REAL-WORLD APIS

Paper: https://arxiv.org/pdf/2307.16789

Github: https://github.com/OpenBMB/ToolBench

9. RestGPT

Paper Name: RestGPT: Connecting Large Language Models with Real-World RESTful APIs

Paper: https://arxiv.org/pdf/2306.06624

Github: https://github.com/Yifan-Song793/RestGPT