AutoGen: A Multi-Agent Framework for Enabling Next-Gen AI Applications - Berkeley 2024

[Music] uh my colleague and me we designed the aogen like we started and released in October uh last year and what was our idea the idea was that uh we wanted to see the future of the AI application especially after have seen such uh kind of amazing power of the llms and uh how can we Empower every developer to build this kind of application basically aogen is programming framework uh for agentic AI so where you have a multi agent framework to solve a a complex task for uh day uh for your daily usage so originally before we even U kind of released the uh G GitHub repo so we uh tested our app our programming framework on uh six different uh available for Academia Frameworks uh such as the mass solving problems uh uh retrieve for gentic uh question answering and decision making is household task that's just available to us uh data set and uh we also tried it for supply and chain optimization and uh also some other uh things we spend a lot of efforts on evalation and by saying we in particular I spend a lot of efforts on evalation and we will uh deeper dive into the evaluation uh uh in the next slides but uh we also tried to enable more people to build agentic AI so that's why we did released the aogen studio which makes it very easy for you to build the workflow even if you don't have uh any coding experience so uh so we working towards learning teaching and optimization so which is again very important topic and uh we need to pursue it uh quite um in um we're trying to pursue it with all the uh community and you have opportunity to uh uh to add to that too and of course uh we want to um uh go through uh different types of Integrations basically the main goals is like we have seen this spark amazing opportunities and cou ilities which llm brought us but can we convert them into the utilities that we actually Mak in the end user life better looking into also like getting the um inspiration from the robotics uh uh Frameworks where uh satisfaction is defined but also uh people look at the different criterias so we designed this very simple uh um framework so it's getting more complicated now if you will find it in the library but stick with me just like this is a a main thing here so you have a cretic edge so this is like uh one of the agents where you describe your task and you give it the logs of your uh Solutions like basically again what kind of logs we have uh we use a conversation programming so those agents are conversing to each other to solve the task so and then uh the agent is able to come up with uh various criterias and it's accepted value for example here I demonstrated for the mass problems because I assume everyone of us are kind of uh advanced in math and can uh see uh this makes a lot of sense there's no effort from you you just give it description to the critic and critic come ups with criteria which might be important for your solution and then the quantifier agent can take any uh solution you proposed and can quantify how is your how is your uh particular solution kind of doing well or not well on a particular criteria by the way it doesn't have to be Corr uh doesn't correlate or may not correlate with uh end satisfaction because it's not there for this it's for you to understand the utilities like is it any utility you bring to the end user and here I wanted to also emphasize that when you uh create those criterias you need to kind of uh think about those as a task based criteria the criteria which is just based on task description but you also need to provide the solutions have you remember we have seen the uh code efficiency here so basically like those one of those kind of criterias would not be suggested for you by llm because it's uh kind of not seen uh before that uh programming languages would be uh used for solving Mass problems how can we uh trust the quantifier because it's just llm call and we also concern how can we trust the quantifier is it robust or not and that's why we are working on towards adding the verifier here where we can uh basically simply uh verify if uh the criteria is not necessarily bad criteria might be good but it just uh data you have there is not enough to assess the craters making we eating the verifi for you which uh basically you can of ront it in one go uh critic quantifier and then the verifier we can suggest you which criteria you should or you should not trust and hopefully return you the score for this and again you could do it with a number of uh depending on like how many calls to LM you want do how slow slow is that but basically without any uh efforts you can Bas uh assess the quality of your application and see what is uh delivered to the end user so and join our Discord and and uh there's number of ready research papers happening on top of the aogen and we try to have our blog uh up to date but most importantly if you think it's important research Direction uh uh help us to develop it and contribute thank you thank you so much