0:00 Introduction and Code Walkthrough 23:54 Basic Agent Structure 32:04 Prompt for Agents 36:13 RAG on Demand Agent 39:08 Grounding Message Agent 46:31 Red-teaming and Blue-teaming Agents 48:26 Advanced Topics - Dynamic Conversation Flow 50:44 More Agents = Better? 53:27 Self-correcting code with a suite of agents 55:00 Pseudo Multi-Agent Conversation 57:45 Is conversation really necessary to do tasks? 59:02 Takeaways from AutoGen 1:03:34 Discussion 1:24:05 Demo by Sarkar!
I have yet to use TaskWeaver. For CrewAI, I think it is one of the better agentic frameworks out that, but it can also be too verbose since it is conversational-based. LangGraph tries to do what TensorFlow did instead of native Python. It feels unnatural to use, I recommend not using it.
My template AutoGen notebook: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/AutoGen/AutoGen.ipynb My AutoGen Slides: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/AutoGen/AutoGen%20Slides.pdf
I think Crew AI is actually easier to use as their crew, task format is quite understandable and intuitive. AutoGen, however, is a little more versatile and can do more complex workflows since they have a GroupChatManager to handle it. I haven't used LangGraph, so I cannot comment. But LangChain agents in general don't have very good prompting, and doesn't work that well based on how I tried them last year. I am in the middle of creating my own StrictJSON agent library, which will use JSON as the root means of communication. I believe having JSON at the root of all agents' output can help save trouble parsing output fields (right now most agentic structures just do regex directly on the text, which can fail if LLM outputs wrongly). Stay tuned.
@@user-wr4yl7tx3w can join my discord. it's in my profile links :) I actively share ideas there - some of my latest ideas don't have papers yet, but the ideas are shared openly
0:00 Introduction and Code Walkthrough
23:54 Basic Agent Structure
32:04 Prompt for Agents
36:13 RAG on Demand Agent
39:08 Grounding Message Agent
46:31 Red-teaming and Blue-teaming Agents
48:26 Advanced Topics - Dynamic Conversation Flow
50:44 More Agents = Better?
53:27 Self-correcting code with a suite of agents
55:00 Pseudo Multi-Agent Conversation
57:45 Is conversation really necessary to do tasks?
59:02 Takeaways from AutoGen
1:03:34 Discussion
1:24:05 Demo by Sarkar!
Thanks for the review, very useful!
great content and thorough walk through. cheers!
One-pizza agent groups sounds like a good idea. It might not just be a limitation of AI and LLMs.
wow this is handsdown the best guide I found joined the discord 🎉
Hierarchical Autonomous Agent Swarm I am going to development with an oversight board if possible in AutoGen. I like the examples given in the video.
John, do you have perspective on other multiagent frameworks like TaskWeaver, CrewAI (based on LC) and LangGraph?
I have yet to use TaskWeaver. For CrewAI, I think it is one of the better agentic frameworks out that, but it can also be too verbose since it is conversational-based.
LangGraph tries to do what TensorFlow did instead of native Python. It feels unnatural to use, I recommend not using it.
Great content!!
Thank you this helps me a lot
For those interested in the StrictJSON framework I talked about at 34:18, here it is: github.com/tanchongmin/strictjson
My template AutoGen notebook: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/AutoGen/AutoGen.ipynb
My AutoGen Slides: github.com/tanchongmin/TensorFlow-Implementations/blob/main/Paper_Reviews/AutoGen/AutoGen%20Slides.pdf
How would you compare it against Crew AI and LangGraph?
I think Crew AI is actually easier to use as their crew, task format is quite understandable and intuitive. AutoGen, however, is a little more versatile and can do more complex workflows since they have a GroupChatManager to handle it.
I haven't used LangGraph, so I cannot comment. But LangChain agents in general don't have very good prompting, and doesn't work that well based on how I tried them last year.
I am in the middle of creating my own StrictJSON agent library, which will use JSON as the root means of communication. I believe having JSON at the root of all agents' output can help save trouble parsing output fields (right now most agentic structures just do regex directly on the text, which can fail if LLM outputs wrongly). Stay tuned.
@@johntanchongmin great! how can I find out more about what you do and your research in Singapore? 👍
@@user-wr4yl7tx3w can join my discord. it's in my profile links :)
I actively share ideas there - some of my latest ideas don't have papers yet, but the ideas are shared openly