All of these tools support the specification of multiagent workflows.
LangGraph takes a graph based approach, specifying the sequence of actions as a graph - nodes and edges (which can be unintuitive at times). CrewAI has many similarities with AutoGen beginning with defining agents and then getting them to address tasks by exchanging messages.
All these frameworks all have the general principle (described in the multiagent book - https://multiagentbook.com/ ) of
- defining an agent,
- giving it access to an LLM, tools, memory
- specifying how the agents communicate within a workflow (e.g., GroupChat in AutoGen, Crew in CrewAI, StateGraph in LangGraph)
I expect our general understanding of these terms will converge as practice continues!
You write so excellently. Currently studying agents for a project I intend working on soon and this has proven very useful in shaping my considerations and planning. Thank you!
Excellent summary of the landscape to date. I agree that the combination and use of various models not only utilizes the best strengths of each one but also creates a team of machines that are dedicated to certain tasks but are integrated into a multimodel environment. The human brain is still the most powerful of all multitasking entities in this world but with the assistance of applications capable of processing data at speeds beyond organic limits the synergy will be pure optimization on every level
The human brain is particularly efficient and has inspired many of the ideas in this space (mixture of experts, pathway models, ... multi agent LLM apps)
It is still an emergent area and more work is done to benchmark and optimize efficiency (efficient orchestration).
Very well articulated. We are trying to use autogen to improve better financial decision making for startups by incorporating inputs from multiple agents! Let's see how things pan out :)
Great write up there Victor, where do you recommend one start from in this large landscape of agents and multi agents?
Thanks for checking out the article. I’m working on a book that is still being written. The first few chapters are available
https://newsletter.victordibia.com/p/announcing-a-new-book-multi-agent
I am considering a shorter course though.
Great post, Victor! It would be interesting to know your thoughts on LangGraph and CrewAI
Great question @Grigory!
All of these tools support the specification of multiagent workflows.
LangGraph takes a graph based approach, specifying the sequence of actions as a graph - nodes and edges (which can be unintuitive at times). CrewAI has many similarities with AutoGen beginning with defining agents and then getting them to address tasks by exchanging messages.
All these frameworks all have the general principle (described in the multiagent book - https://multiagentbook.com/ ) of
- defining an agent,
- giving it access to an LLM, tools, memory
- specifying how the agents communicate within a workflow (e.g., GroupChat in AutoGen, Crew in CrewAI, StateGraph in LangGraph)
I expect our general understanding of these terms will converge as practice continues!
AI agents will become the "apps" of the era of large language models, exploring new paradigms of human-computer interaction and collaboration.
LLMs have some known drawbacks:
- They can generate hallucinations.
- The results are not always true.
- They have limited or no knowledge of current events.
- They struggle with complex calculations.
- They lack the ability to take actions.
- They lack long-term memory capabilities.
Agents can utilize external tools to overcome these limitations. For example:
- Search engines: to obtain the latest information
- Python REPL: to execute code
- Wolfram: to perform complex calculations
- External APIs: to access specific information
You write so excellently. Currently studying agents for a project I intend working on soon and this has proven very useful in shaping my considerations and planning. Thank you!
Glad you found it useful. Its an interesting space.
Feel free to use and cite.
Excellent summary of the landscape to date. I agree that the combination and use of various models not only utilizes the best strengths of each one but also creates a team of machines that are dedicated to certain tasks but are integrated into a multimodel environment. The human brain is still the most powerful of all multitasking entities in this world but with the assistance of applications capable of processing data at speeds beyond organic limits the synergy will be pure optimization on every level
Agreed.
The human brain is particularly efficient and has inspired many of the ideas in this space (mixture of experts, pathway models, ... multi agent LLM apps)
It is still an emergent area and more work is done to benchmark and optimize efficiency (efficient orchestration).
Very well articulated. We are trying to use autogen to improve better financial decision making for startups by incorporating inputs from multiple agents! Let's see how things pan out :)
Great! I'd love to learn more about what you find!
i think next big thing will be algorithm to combine and merge different llms efficiently, to reuse all training