Nice post! Multiagent Protocols is indeed a good theme to think about in the context of Agentic. Just wanted to say, however, that protocols is a more general theme and that there has been tremendous progress in modeling, verifying, and implementing multiagent interaction protocols. You can find some software here: https://gitlab.com/masr/
Nice of you to ask! Reading your post again, if I had to add something, it would probably be along something along these lines:
Every application is different, which would motivate a different protocol between its agents. One standard protocol is not going to fit the bill. We need approaches for modeling and implementing interaction protocols that are based on standards (protocol language, wire formats, and so on). This would simplify implementing the communication-related aspects of agents (which are often the most complex pieces in loosely-coupled systems) and promote interoperability (agents could be constructed by independent parties based on the protocol).
That's where the work we are doing comes in. Traditional protocol approaches model protocols in terms of low-level state machines (much like UML interaction diagrams) that capture message ordering. Such approaches do not lend themselves to modeling flexible interactions, which is a huge drawback for AI agents. Our approach, based on information protocols (https://www.csc2.ncsu.edu/faculty/mpsingh/papers/mas/AAMAS-11-IBIOP.pdf), supports maximal flexibility because instead of specifying a protocol in terms of message ordering, it specifies them in terms of information dependencies: An agent can send a message when it has the necessary information (which may have come from any agent in the system).
Nice post! Multiagent Protocols is indeed a good theme to think about in the context of Agentic. Just wanted to say, however, that protocols is a more general theme and that there has been tremendous progress in modeling, verifying, and implementing multiagent interaction protocols. You can find some software here: https://gitlab.com/masr/
and papers about the software here: https://www.lancaster.ac.uk/staff/chopraak/
We are currently extending the tool suite in many ways. Happy to chat further!
Best
Amit Chopra
Thanks Amit, would love to learn more about the masr project. Will checkout the repo. What specifically would you add to, or update in the blog post .
Nice of you to ask! Reading your post again, if I had to add something, it would probably be along something along these lines:
Every application is different, which would motivate a different protocol between its agents. One standard protocol is not going to fit the bill. We need approaches for modeling and implementing interaction protocols that are based on standards (protocol language, wire formats, and so on). This would simplify implementing the communication-related aspects of agents (which are often the most complex pieces in loosely-coupled systems) and promote interoperability (agents could be constructed by independent parties based on the protocol).
That's where the work we are doing comes in. Traditional protocol approaches model protocols in terms of low-level state machines (much like UML interaction diagrams) that capture message ordering. Such approaches do not lend themselves to modeling flexible interactions, which is a huge drawback for AI agents. Our approach, based on information protocols (https://www.csc2.ncsu.edu/faculty/mpsingh/papers/mas/AAMAS-11-IBIOP.pdf), supports maximal flexibility because instead of specifying a protocol in terms of message ordering, it specifies them in terms of information dependencies: An agent can send a message when it has the necessary information (which may have come from any agent in the system).
This Blue Sky paper gives the history of communication in multiagent systems and will likely be breezy reading for you: https://www.lancaster.ac.uk/staff/chopraak/pdfs/MOS_Blue_Sky.pdf
Thanks for the detailed response, Amit. I will read the two papers you've linked above before responding.
You should connect with Ockam. They've built the hard parts of this. Reach out to Mrinal their CTO.
Thanks Matthew, not sure if I've come across Ockam before. Do you have their website or github url?
https://www.ockam.io/