Agents

•

Introduction to AI Agents, Langchain, SWE.

Agents

LLM Agent Paper List

Paper: The Rise and Potential of Large Language Model Based Agents: A Survey

Survey of LLM-based AI Agents

Paper: An In-depth Survey of Large Language Model-based Artificial Intelligence Agents

Blog: 4 Autonomous AI Agents you need to know

Camel

Communicative Agents for “Mind” Exploration of Large Language Model Society
Code: https://github.com/camel-ai/camel

AutoGPT

Paper: Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions
Code: https://github.com/Significant-Gravitas/Auto-GPT
Blog: AutoGPT architecture & breakdown

Tutorials: AutoGPT Forge

AgentGPT

Code: https://github.com/reworkd/AgentGPT

BabyAGI

Blog: Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications

Colab:

Godmode

Voyager

Paper: Voyager: An Open-Ended Embodied Agent with Large Language Models
Code: https://github.com/MineDojo/Voyager

Talk2Drive

Paper: Large Language Models for Autonomous Driving: Real-World Experiments

MemGPT

Paper: MemGPT: Towards LLMs as Operating Systems
Code: https://github.com/cpacker/MemGPT

RL-GPT

Paper: RL-GPT: Integrating Reinforcement Learning and Code-as-policy

MC-Planner

Paper: https://arxiv.org/abs/2302.01560
Code: https://github.com/CraftJarvis/MC-Planner

Multi-Agent

Generative Agents

Paper: Generative Agents: Interactive Simulacra of Human Behavior«br> Code: https://github.com/joonspk-research/generative_agents
Demo

Blog: Paper Review: Generative Agents: Interactive Simulacra of Human Behavior

Multi-Agent Collaboration

Multi-Agent examples

LangGraph + Llama3 + Groq

Colab: https://drp.li/X3hpZ

Frameworks

LangChain

LangGraph

Introduction to LangGraph

SWE

OpenDevin

Paper: SWE-AGENT: AGENT-COMPUTER INTERFACES ENABLE AUTOMATED SOFTWARE ENGINEERING
Code: https://github.com/OpenDevin/OpenDevin
Docs: OpenDevin Intro

DroidAgent

Paper: Autonomous Large Language Model Agents Enabling Intent-Driven Mobile GUI Testing
Code: DroidAgent: Intent-Driven Android GUI Testing with LLM Agents

WebVoyager

Paper: WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models
Code: https://github.com/MinorJerry/WebVoyager

AutoCodeRover

Paper: AutoCodeRover: Autonomous Program Improvement

ReAct

Chain of thought and ReAct

  1. Thought: The reasoning step, or thought, serves as a guide to the Foundation model, demonstrating how to approach a problem. It involves formulating a sequence of questions that lead the model to the desired solution.
  2. Action: Once the thought is established, the next step is to define an action for the Foundation model to take. This action typically involves invoking an API from a predefined set, allowing the model to interact with external resources.
  3. Observation: Following the action, the model observes and analyzes the results. The observations become crucial input for further reasoning and decision-making.

List chain-of-thought steps:

List ReAct steps:

Langchain ReAct

Llamaindex ReAct

ReAct Agent - A Simple Intro with Calculator Tools
ReAct Agent with Query Engine (RAG) Tools
Controlling Agent Reasoning Loop with Return Direct Tools

Fine-tuning a gpt-3.5 ReAct Agent on Better Chain of Thought Custom Cohere Reranker

Octpus

Paper: Octopus: Embodied Vision-Language Programmer from Environmental Feedback
Code: https://github.com/dongyh20/Octopus

Octpus v2

Paper: Octopus v2: On-device language model for super agent
model: NexaAIDev/Octopus-v2
kaggle: https://www.kaggle.com/code/rkuo2000/octopus-v2

Octpus v3

Paper: Octopus v3: Technical Report for On-device Sub-billion Multimodal AI Agent

Octpus v4

Paper: Octopus v4: Graph of language models
model: NexaAIDev/octo-net
Code: https://github.com/NexaAI/octopus-v4

ADAS

Paper: Automated Design of Agentic Systems
Code: https://github.com/ShengranHu/ADAS

Gödel Agent

Paper: Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement
Code: https://github.com/semanser/codel
語言模型隨著代理系統(agentic system)的發展，在推理、工作規畫等領域有很大幅度的進步，這些代理系統(agentic system)主要可以分為兩大類，一是固定整個工作流與工作模組的Hand-Designed Agent；另外一種則是允許較彈性的工作流，並讓Agent可以適度選用工具的Meta-Learning Optimized Agents。但這兩者皆是基於人類先驗經驗而設計的系統，它將受限於人類的經驗，使得整個系統失去最佳化的可能性。
本篇論文的研究團隊，嘗試利用哥德爾機(Gödel machine)的概念，讓Agent可以自行決定工作流程，自行選用工具模組，並依照環境反饋，自我改良整個工作系統，研究團隊將其稱為Gödel Agent。

在這篇概念性的論文中，研究團隊指出如果要能實現自我優化，Gödel Agent至少需具備四種能力：

  1. 自我覺察(Self-Awareness) 能夠讀取工作流當下，環境與Agent的各式變數、函式、類等參數值，取得整體的運作狀態(operating state)。
  2. 自我改善(Self-Improvement)
  3. 能夠利用推理與規劃的能力，針對當下狀況，認知到應該調整那些工作區塊，並進而調整程式碼去修改工作邏輯。
  4. 與環境互動(Environmental Interaction) 針對當前修改的結果，可以由環境的狀態變化取得反饋，得知目前的策略是否成功，並評估是否需要再度調整。
  5. 持續改進(Recursive Improvement) 利用前三項能力，不停迭代，在經過幾次迭代後，這將產生類似Gödel machine的效果，以達到整個系統的最佳化。

研究團隊給予Gödel Agent幾種不同的工作類型，測試它的表現，並與過往幾種方法，諸如CoT、Self-Refine、Role Assignment、Meta Agent Search進行比較，就結果上來說Gödel Agent完勝，不過由於測試的工作類型較侷限，目前尚未知道Gödel Agent在不同的領域是否都如此出色。
研究團隊也給出未來的研究方向，例如語言模型能否產生集體智慧(collective Intelligence)，或者Gödel Agent是否確實達到系統理論上的最佳化(theoretical optimality)等，都是有趣的研究主題。

OpenAI Swarm

llama3-groq

import openai
from google.colab import userdata

model = "llama3-groq-70b-8192-tool-use-preview"

llm_client = openai.OpenAI(
  base_url="https://api.groq.com/openai/v1",
  api_key=userdata.get('GROQ_API_KEY'),
)

bare_minimum

# https://github.com/openai/swarm/blob/main/examples/basic/bare_minimum.py
from swarm import Swarm, Agent

swarm_client = Swarm(client=llm_client)

agent = Agent(
    name="Agent",
    instructions="You are a helpful agent.",
    model=model,
    tool_choice="auto"
)

messages = [{"role": "user", "content": "Hi!"}]
response = swarm_client.run(agent=agent, messages=messages)

print(response.messages[-1]["content"])

Swarm_Llama3-Groq.ipynb

MLE-bench

Paper: MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering
Code: https://github.com/openai/mle-bench



This site was last updated November 15, 2024.