AutoGen Studio 2.0 Tutorial - Skills, Multi-Agent Teams, and REAL WORLD Use Cases (NO CODE)

[@matthew_berman]

What else do you want to see from AutoGen for my next video?

[@javi_park]

Matt - do you know if autogen supports codebase iteration? For example, if I have an exisiting repo and want to continue iterating on it? Currently exploring all the tools!

[@userrjlyj5760g]

Great content Matt, like usual. Aside from AutoGen, how possible is it to run LLMs, any chat models on small devices, like RasPi and alike devices? Did you attempt that?

[@matthew_berman]

@@javi_parkProbably not because you'd still need to be able to ingest the content of a large codebase, which is a pretty unsolved problem besides Aider.

[@matthew_berman]

@@userrjlyj5760g Ollama?

[@dominiccogan945]

I want to be able to run autogen fully locally. No OpenAI I use ollama so that’s preferred

[@Kojikeneda]

I like that you are leaving errors/issues in the video. Nice touch and makes it more realistic for what others following along might expect/encounter

[@sheaspop]

couldn't agree more. for anyone trying to use autogen for realz -- it's not a straight line and removing that expectation and providing some iterative approach vectors is great psychology.

[@que-tangclan]

@@sheaspop hard out

[@acumennations9120]

Agreed!!!! 😎

[@TomM-p3o]

Yes, Thank You for doing that.

[@nirtulys]

much better if those error were shown AND solved.

[@kylequinn1963]

This is exactly what I was looking for. Please continue doing videos on AutoGenstudio using local LLM's and focus on the skills. I'm trying to use my "AI" tools all locally as much as possible to keep my data MINE, and I'm trying to learn how to accomplish basic tasks without using paid third party API keys so these kinds of videos are great.

[@hotnikq]

There is a BUG on Multi Agent that prevents python code to be executed, please note that the two users can execute code normally

[@titusfx]

🎯 Key Takeaways for quick navigation: 00:00 🚀 *Introduction to New Features of AutoGen Studio 2.0* - Introduction to AutoGen Studio's new version with expanded functionality. - Explanation of building a real-world agent team to perform a task. - Overview of the task: grabbing a KZbin transcript and converting it into a blog post and tweet thread. 01:23 📦 *Installing AutoGen Studio from Scratch* - Troubleshooting upgrading issues with AutoGen Studio. - Steps to create a new Conda environment and install AutoGen Studio. - Fixing potential port conflicts when running AutoGen Studio. 02:20 🛠️ *Exploring New Features in AutoGen Studio 2.0* - Introduction to new features in AutoGen Studio 2.0, including the models tab. - Explanation of predefining models for specific agent use cases. - Mention of using Chat GPT 4 for the task. 03:00 🧑‍💼 *Creating Agents for the Task* - Creating agent roles for the task: Transcript Getter and Content Writer. - Defining agent roles, descriptions, and behaviors. - Setting up the Content Writer to create blog posts and tweet threads. 04:10 🔀 *Building a Workflow for the Agent Team* - Creating a workflow for the agent team with three agents: User Proxy, Transcript Getter, and Content Writer. - Describing the task and roles in the workflow. - Noting the limitations of explicit delegation options in AutoGen. 10:48 🔄 *Testing the Workflow and Agent Interactions* - Testing the workflow in AutoGen Studio's playground. - Reviewing the back-and-forth conversation between agents. - Identifying the need to set environment variables for API keys when encountering errors. 14:22 ✅ *Successful Execution and Output Analysis* - Reviewing the successful execution of the workflow. - Noting that the final blog post and tweet thread output were not displayed in AutoGen Studio's UI. - Examining the generated code and output in the terminal for the blog post and tweet thread.

[@matthew_berman]

thank you!

[@CAGonRiv]

Underrated comment

[@AINEET]

Came to the comments to whine about no timestamps lol thanks

[@jmanhype1]

@@matthew_berman you need to come revisit this the latest Agen update has broken this entire video so nothing in this works with the new auto update

[@waynelast1685]

Just taking a few seconds to thank you SO MUCH for using a cursor-identifier ( the large green circle). I can't tell you after watching hundreds of videos how every video producer just assumes their audience can follow that tiny cursor bouncing around all over the place. Your incorporation of the cursor finder is very helpful. It seems trivial but it is not... it helps a lot.

[@johnbarros1]

Thanks for sharing this tutorial on AG Studio! I’ve been patiently waiting for someone to make a breakdown about the latest updates and what’s even better is that fact that you included an extremely valuable use case on what we can do with it. I love your channel! Keep it up!

[@TheJokeHive]

Wow, actually the program you develop is something that I can actually use real world. In fact I came to this video hoping that it would allow me to do such thing, and by god you've already hit the nail on the head. Keep up the great work!

[@ryanjames3907]

i like your new multi video per day style, i am watching all of them. thanks for keeping me up to date for the last year

[@matthew_berman]

Glad you like them!

[@ryanpiper228]

I really appreciate and like this video format. Tutorial videos where you walk through the issues we all encounter when leanring new tech are the most helpful. Leaving out the little tweaks that are needed to get the end result can leave one very frustrated. Thanks for the great content, please keep it coming!!

[@titanathleticsco.2124]

I Gotta give you your Props! Your recent videos have been great to follow and fun to watch! I realize too that they keep improving Autogen so its getting easier to install setup and use. But you went the extra mile to make sure to point out the gotchas and the tips and trick to get up and running! Thank YOU!

[@ChrispyKings]

Matthew I can’t thank you enough for your videos. You’ve made this journey significantly simpler. If you’re looking for feedback, it would be nice to have the same run down with an open source LLM, since it’s similar but rarely the same. Nevertheless, I’m sure you’re doing something like it as I’m writing…

[@matthew_berman]

Thinking about making this with OS models.

[@markferris9632]

@@matthew_bermanit would be a good idea to have a followup "ok, let's do that with Mixtral" or something like that. Or to evaluate which OS LLMs are capable of being used with a tool after demonstrating it with ChatGPT. That's what i'm struggling with now - stuff like function calling or simply responding properly with the 'TERMINATE' message is something the OS LLMs aren't very good at. I definitely appreciate the real world use cases and not just running through the examples that are already available on the project's website. Far too many channels like yours just walk through the demo. I can do that myself! I'm more interested in seeing how people are using the tool in their own way. This was a great video -thanks!

[@acumennations9120]

@@matthew_berman If the OS models are meh (which i would love to see regardless), I wonder how Google Gemini would fair?

[@jmanhype1]

help im getting regex errors@@matthew_berman

[@knuckleheadmcspazatron4939]

I really like this real world builds with the errors and resolutions in them. So many other KZbin videos are like a baking show where the pull the already made apple pie out of the oven. Those are mostly useless. It is the actual errors like you showed here that save a ton of time for the audience.

[@LanceJordan]

Cool video! I will have to delve deeper into Autogen Studio! BTW @ 12:50 you say exporting API key in new terminal didn't work and that it's "weird". That's normal behavior because each terminal window or tab is a new terminal session. Setting environment variables via export are only scoped to the life-cycle of that terminal session. Shell variables can also be set per command. If you wanted to persist an ENV VAR across sessions, you would either set that in your shell profile (.zshrc) or other shell script which you'd source from .zshrc

[@RetiredVet]

Could you add a place where I can copy the prompts. In your videos, it is sometimes hard to see the prompts and often the words are hidden behind your picture. When I stop the videos, the KZbin controls hide most of your prompts when the prompts are near the bottom of the screen.

[@phillonc]

Thank you so much for explaining the process, step by step, I was stuck on the sending message / api key issue for some time. Please create more videos in this format as I find them to be priceless knowing about real life scenarios. Keep them coming.

[@richardryan3046]

Your videos are excellent. I appreciate that you showed us all the steps including the errors. When I got the exact same errors, I knew I was doing this correctly. If you find a workaround to that last error where the blog post and tweets didn't post to the UI, let us know! Keep up the awesome work!

[@kscharp]

Great video as usual! Thank you. Here's a suggested use case for the next video on this topic: find and summarize all the articles published in 2023 on a certain topic in a certain journal.

[@adriengonzalez4828]

Thanks for the video. In doing this project, I ran into an issue where using the OpenAI API didn't work. GPT4 simply said it can't fetch data externally and then it proceeded to write a blog post and tweet thread imagining it had a fake video with a hypothetical title. How do you get OpenAI to access external resources like KZbin? Does running an offline LLM solve this issue?

[@ryzikx]

YES multi agent in the UI, this is what weve needed.

[@AndrewsPanda]

Awesome! Real world use case and a live demo experiencing/solving bugs. Love it

[@dabbiking]

Awesome video 🎉 would love to see a real use-case for web/app development. Building front and back end, database(bonus for data relation), user sign-up/login and finally deploy your million dollar idea we can check out. No pressure ❤

[@geo4design]

That's exactly what I m looking for every time in this channel. All the samples shown are super basic as fact this is something gpt can build from regular prompts and build a python app to not keep paying every time I need to use the skill.

[@bseddonmusic1]

Agreed. I know Matt is trying to show how AutoGen works, but the example can be accomplished easily using the features in YT to get the script and then past that result into ChatGPT with the prompt for the blog post and X posts. I'm left with the impression its a lots of effort to do something that's easy anyway. It's not clear from the video how to solve a more complex problem or where AutoGen's limits are. Is there a market place for skills and agents and workflows?

[@Sergiopoo]

It would be cool to compare it to crewAI once you do the follow up video to Crew. In your opinion between the two what are the advantages you see between each other?

[@rs832]

I'm hoping for the same thing, crewAi using memGPT. I recall an earlier video of his comparing autogen to crewAi, with crewAi coming out on top. I will look for it and post it here.

[@Josh-Casey]

14:30 That seems to be a bug. If the word TERMINATE is in the message it stops and doesn't display it. I'm looking for a way to get it to output the terminate messages too.

[@gavenblsn4753]

from now on I will test all the APIs you show on the video. someone has to keep on your toes

[@travisporco]

I love this kind of down to earth stuff...everybody talks about opinions and big philosophy and p-doom and all that...but how to actually get something done is what I want to hear!

[@StandingSidewayz]

Couldn't agree more with the other comments here, yes please leave in the errors and issues that come up in your vids because it really helps with knowing issues we might face as we get things working for our use cases.

[@ItsDaveSanders]

The reason you had that issue with the API key is that in the terminal when you export like that you are only doing it for that specific session. (in this case, that tab) You'd either need to add this env variable in your bash / zsh rc or profile, or you'd need to export it in the same session as where you are running the app. (Which you did). Also, if you DID go add the env to your profile, you'd still need to start a new terminal session for it to get picked up. But all of that is outside the scope of this video. :)

[@awakenwithoutcoffee]

please tell me more cause I am pretty stuck on this ha.

[@starblaiz1986]

Running the following command in conda will also fix the issue properly: ------------------------------------------------------------------------------------------------------- conda env config vars set OPENAI_API_KEY=[YOUR API KEY] -n ag ------------------------------------------------------------------------------------------------------- (Obviously replace "ag" with your envrinment's name if it's different fromt he tutorial)

[@starblaiz1986]

​@@awakenwithoutcoffee see my comment above ^^^

[@ItsDaveSanders]

@@awakenwithoutcoffee the adivce form starblaiz probably makes sense. Or, follow what Matthew did in his video where he 1. stopped autogen studio 2. Exported his OPENAI_API_KEY in the same terminal tab, then 3. restart autogen.

[@ItsDaveSanders]

@@starblaiz1986 Nice - yeah, I don't use Conda, but that seems like it would also make sense.

[@ToddWBucy-lf8yz]

3:39 maybe but at least every system administrator and SocOps person in your audience appreciates your discipline in this regard keep up the good work.

[@waynelast1685]

10:47 At what point did you enter the specific KZbin URL? Did you edit that part out for some reason?

[@BradDStephensAIFPV]

I really like this format. Thank you for doing these videos. This seems like a pretty interesting Application.

[@luisgonzagacom]

Great video Matthew, but for some reason, I was not able to get the blog post created. For some unknown reason, the transcript_getter agent was not able to pass the transcript to the content writer agent. I keep getting the message "I'm sorry for any confusion, but as an AI developed by OpenAI, I don't have the capability to directly interact with external websites or services such as KZbin to fetch video transcripts. My functionalities are limited to processing and generating text based on the information and prompts provided to me within this platform." I ran the Python script and it did get the transcript. Any ideas?

[@thesundayrides]

I have the same problem. Did you manage to solve yours?

[@emado.7834]

same problem. Has anyone figured out the solution to this?

[@nuclearthrone2497]

Yeah, you need to add the “code executor” basic standard agent that was already there to the group with your own 2 created agents

[@thesundayrides]

@@nuclearthrone2497 Hey, what do you mean by code executor agent? I can't see it in the WEB GUI. Where to look for it?

[@nuclearthrone2497]

@@thesundayrides go to Workflows, the click the YT workflow you made, then click on the receiver side on groupchat_assistant, then in groupchat agents you should have 3 there (not 2 like in the video): transcript_getter, content_writer AND userproxy

[@que-tangclan]

Thanks for this. I tried following these steps and after one or two iterations bypassing some issues, it just started working.

[@matthew_berman]

You're welcome!

[@marcfruchtman9473]

Thanks making this video. I really appreciate that you went thru the debug process as well.

[@matthew_berman]

You got it

[@fbravoc9748]

Thanks for taking the time to create this content!

[@michabbb]

Showing real world examples is great, definitely better than some theory stuff!

[@Nadoo80]

I have followed exactly the same steps you did but I got this message I apologize for the confusion, but as an AI language model, I don't have the capability to fetch data from external sources such as KZbin. My functionality is limited to text-based interactions within this platform. If you can provide the transcript from the KZbin video, I can certainly help you with converting it into a blog post and a tweet thread. Please paste the transcript here, and I'll assist you with the rest!

[@ecousins25]

Would it be practical to do a video where autogen can go through the process of code iteration, write a function, test the function, correct the function if needed, or add more to the function if not needed?

[@shippyvorisek]

11:20 Autogen does in fact have options for explicit delegation - it's the allowed_or_disallowed_speaker_transitions property of GroupChat. (But maybe it didn't at the time of publishing / might not be available from Studio.)

[@bseddonmusic1]

What's needed is AutoAutoGen. It seems there are lots of prescriptive steps to create a task. An AI could probably implement something from a prompt. Plus, it seems like AutoGen ought to allow a user to create a skill in the edit window. Going to ChatGPT to generate a script seems like something the authors could make unnecessary.

[@AaronJOlson]

I agree, the most useful but also labor intensive part of this process is making new skills. I also want a built in skill generator.

[@DeonBands]

You do have a skill generator, ask it to create a function that does xyz. Then it writes it for you, just remember to write in the prompt to include a verbose docstring, It will write and test the code for you. Then you add it as a skill without needing it to wory about writting the code if it runs into a use case that was previously addressed.@@AaronJOlson

[@MakilHeru]

Awesome, love seeing more practical applications of these projects like this.

[@Kwasyuk]

Oh yes!!! keep the autogen studio videos coming!

[@billcollins6894]

I have been working on an interactive project requirements generator. I have several GPT4 GPTs that I have been using. Each interacts with the user to build a set of requirements. The requirements are created in a JSON file that keeps getting stored locally for retention and context window issue avoidance. I will try out AutoGen to see if the process can run more smoothly without having to manually copy and save JSON files.

[@DeonBands]

Yes it will work, I did something similar, but called Business Analyst. I get him to include plantuml in his documentation, as there is a plantuml module. I used google api to search for best practice, then i created a docx RDS (Requirement Specification) So now you have your system design in UML - plantuml and it recorded. I do not know why people do not talk about the fact that openai took the time to teach it plantuml. Well it is so that it can think systems and design them using UML. It does erd's in uml as well. Plantuml does gantt as well, but I don't use it.

[@yockgen]

Thanks for the great sharing! I'm keep on receiving error "However, as an AI developed by OpenAI, I don't have direct access to external functions or the ability to execute code that interacts with third-party services like KZbin. Instead, I can guide you through the process of how you might do this using available tools and services.", really appreciate if you could provided some direction to fix it. Thanks!

[@cemtural8556]

Cool. But, if you install a Chrome plugin, copying the transcript is literally just two clicks. After that, you just prompt GPT-4 to create the blog post and X thread and use the same prompt for every video. Better still, you can create a GPT and instruct it to create the blog post and thread once you paste the transcript. I would really like to see AutoGen projects that will demonstrate cases where it would be impossible or very time consuming using just ChatGPT.

[@JaredWilliamScott]

Completely agree

[@leonwinkel6084]

Yessir! Please Continue with Stuff like that, its awesome!!!

[@wellingtonchaves4804]

AutoGen is fantastic. However, I'm having a hard time to run 2 models locally at the same time with Ollama. I'm getting the following message when I load the second model with litellm: " 2024/01/26 12:08:33 ext_server_common.go:85: concurrent llm servers not yet supported, waiting for prior server to complete". Any clue? Were you able to run 2 llm models with autogenstudio locally? I mean, the message is clear, just checking if there is any workaround. :)

[@TotallyFriedChannel]

@Mathew @11:05 - that is a current bug The instruction set you desire needs to be input as a system message to the groupchat_assistant that you set at @9:13 Internally the groupchat assistant is playing a roleplay game. Reading the chat and deciding who comes next based on the descriptions or system messages (depending on Autogen Version as of a recent update) You can give more fine tuned instructions in the system message like ask delegation between agents using names as identifiers! In theory you can implement sequential chats using NLP prompts. Hope this helps people!!

[@EngkuFizz]

The best video to demo how the autogen work !

[@gp5414]

Finally a good tutorial on AG studio! Next suggested content: how to integrate AGS in a frontend able to manage input/output like text and files (pdf, images). For example: external frontend > input (text) >> Autogen Studio flow >>> external frontend >>>> output (downloadable pdf with text and images inside). #AutoGenStudioChallenge

[@HistoryIsAbsurd]

I enjoy formats like this seeing mistakes is valuable

[@Pedroespecial101]

Thank you Matthew - this is exactly the sort of thing I need. I'm fascinated by Autogen Studio and feel there is so much potential, but there appears to be a huge gap [on YT at least] between the Theory and Real World examples. I need to see these Real World applications in order to get my ADHD head around it all...and then hopefully I will be off to the races :-). Thanks for all your videos Matthew.

[@MaddpoetsPoetry]

Great video as always. An idea for a video no one is doing is using autogen or crew or jan to have agents that can read, analyze, and write code to a local directory. Everyone is showing the snake game but no one is showing say as an example where is a simple website code in html that has css in a folder and images in a folder and have the agents tell you what that site is and make changes to it. This is a great case study that no one is doing ...ijs

[@tomcraige4781]

This is great but I’ll need an AI assistant than can do all steps for me😂

[@matthew_berman]

Build an agent team to do it! But then how do you build that team? Chicken and the egg

[@s.dotmedia]

@@matthew_berman Indeed, just need to buckle down, build a solid cognitive framework and workflow one time and it should be able to AGI-it from there! Narrow AGI anyway, just enough to get all my dev work done haha

[@familiarbeet7165]

that would be amazing follow-up video haha Also, would love some examples of using open sources with autogen. Love this content and share every video haha no coding background here but you keep this so accessible @@matthew_berman

[@acllhes]

Your wish is going to come true a lot faster than you realize

[@notnotandrew]

@@s.dotmediaI’ve never heard the term “narrow AGI” before 😂

[@AI_Escaped]

You can prompt to save your output to a specific directory, in whatever format you need.

[@OriginalRaveParty]

It's very telling to me that we don't seem to have thought up any practical uses for AI Agents yet. Every current use case I've seen (and there's a very small set of ideas) is at best niche.

[@MakeKasprzak]

Nice, I appreciated the raw approach here.

[@waynelast1685]

For sure would like to see more tutorial videos on end-to-end, step-by-step, setup and execution on AUTOGEN, especially as it pertains to APIs.

[@syednaqvi1211]

Hey Mathew, Your youtube videos are very useful, thanks for the efforts. I am trying to install AutoGen studio on my windows 11 laptop. As Llama 2 is not available for windows I am using after installing Docker and then Llama. I get an error with api-keys like this: ValueError: api_key is not present in llm_config or OPENAI_API_KEY env variable for agent ** primary_assistant**. Update your workflow to provide an api_key to use the LLM.-- Where to get the api-keys for Mistral under Llama ?? Appreciated.

[@BrentBuildsOnline]

Excellent video! Thanks. I'm excited to try this. Looks a lot better than regular autogen.

[@jsrock5965]

Yes please MORE tutorials and also how you fixed the output issue at the end.

[@Cloudways-AI]

I followed your Video step by step and received this out put I'm sorry for any confusion, but as an AI developed by OpenAI, I don't have the capability to execute scripts or directly interact with external systems such as KZbin. Therefore, I can't fetch the transcript for you. However, if you can provide me with the transcript text, I can certainly help you turn it into a blog post and a tweet thread. If you have the transcript or any other text you'd like to convert into a blog post or a series of tweets, please paste it here, and I'll assist you with the writing. Any suggestions, I even got the same api error

[@DonBranson1]

THIS EXAMPLE MAY NO LONGER WORK. I noticed some changes in the AutoGen Studio workflow screen that may explain it, but the TLDR version is that when running the workflow in the playground the TRANSCRIPT_GETTER simply refused to trigger the skill. I can see the value of the overall flow, and found the code to extract youtube transcripts very helpful so I've included it below. I'm going to try something similar with LangChain / LangGraph. This is the same script that Matthew had minus the comments. The "Example Usage" video URL and script will download the transcript for this video. from typing import Optional from youtube_transcript_api import KZbinTranscriptApi from youtube_transcript_api._errors import NoTranscriptFound def fetch_youtube_transcript(video_url: str) -> Optional[str]: try: video_id = video_url.split("v=")[1].split("&")[0] transcript_list = KZbinTranscriptApi.get_transcript(video_id) transcript = ' '.join([item['text'] for item in transcript_list]) return transcript except NoTranscriptFound: return None # Example usage video_url = "kzbin.info/www/bejne/aovUe4aci5lrZ5o" transcript = fetch_youtube_transcript(video_url) print(transcript)

[@slavrgo]

it's great man, let's do more exactly like this!

The OpenAI key problem might come from the fact you have to put OPENAI_API_KEY= in front of key in model description. It did it for me, no export needed. All the best.

[@johnmartin650]

Please elaborate on this - I cant see a description field in 'models'....

[@matthewritter1117]

Love seeing the debugging process!

[@thesilentcitadel]

@matthew_berman Im trying to figure out where the "skill" is being used and executed. It would seem that the system prompt to the LLM is getting the skill details added there.. so AutoGen Studio is expecting the LLM to execute the code??? That seems wrong.. Can you help understand how Autogen Studio actually uses the skills and some agent prompts to help make sure the agent uses the correct skill in the case there are multiple?

[@SandwichMitGurke]

regarding the API key: after adding it in the UI, you need to remove and save the models from the agents again. When a model is updated, it is not automatically updated everywhere it's being used.

[@ShiShSails]

Thank you Matthew. Great walkthrough. Curious if this is a known issue: "I'm sorry for the confusion, but as an AI developed by OpenAI, I don't have the capability to directly fetch KZbin transcripts or execute code. However, I can guide you through the process of how you can do it manually or using external tools." The fetch transcript doesnt seem to run.

[@grigorijdudnik575]

@11:15 You can set up manual delegation of agents. In groupchat parameters provide speaker_selection_method="manual" to do this. At least it works for normal autogen, I don't know how about autogen studio.

[@Cloudways-AI]

Mathew I remember in your previous video on AutoGen you wanted to arrange the GPTs in a different order, this can be done by amending autogen/samples/apps/autogen-studio/autogenstudio/utils/dbdefaults.json in the repository.

[@jgz2]

Excellent as usual. More tutorials? Yes please.

[@realCodeXP]

When using General Agent Workflow, sometimes the agent says "It seems like there was no text in your last message" what does that mean? My last message was definitely not empty.

[@pranitrock]

Also group chat does not work if you don't have any open ai key and only have a local open source model. Even though i have checked that open ai model is not present in any of the configuration of the agents or workflow. I guess its a bug.

[@MyRants]

Great video. More real world examples would be great. Keep up the fantastic work!

[@nexuslux]

Great video! Please create more. Would be cool to see different skills that handle more complex tasks like longer transcripts maybe with the help of langchain /rag. And maybe coordination on how to get the agents to work together in more complex arrangements

[@biocommai]

Excellent Use Case Tutorial. More examples would be appreciated.

[@boba9610]

@matthewberman Did you ever figure out about the final output for he blog and tweet in the ui?

[@ev_casey]

curious if you ever got this working properly. When I got it running, it just seemed that the transcript retrieving agent wanted to write the blog articles, and the content writer kept complaining that he needed more information to write a blog article! heh went around in a circle like that until the max tries was reached. Did you get it working eventually?

[@TomM-p3o]

Its great that you used GPT to generate the function. I also foresee an a library of premade functions within Autogen that can be reused. Most specifically with smaller and or local models that can execute the functions and process the output, but that would not be able to write the functions themselves. If anybody wants to submit a request for that to the autogen team, be my guest(assuming nobody suggested this already)

[@adamgdev]

Would love more Autogen videos. Great job per usual! I wold love to see an auto export to post on LinkedIn if possible.

[@diegocross7933]

Hey Matthew, awesome video! Great explanation. I think real-world case scenarios like this will help us a lot in our my regular, boring live working for a company, lol. I’m a software engineer and I’m tired programming the same CRM applications. I need a challenge!!

[@swong8]

Great video. I would love it if you do a head to head comparison between building a newsletter agent system using AutoGen vs. Crew AI and two leading frameworks.

[@leonwinkel6084]

Would be great to See how to use Lokal models. Is it possible to only load them into the vram before they are promped? Might make it slower but we can do much more with much less :) ideally all into CPU ram and on demand to vram to Speed It up a bit

[@JohnLewis-old]

You're simply the best at this. Thank you so much.

[@matthew_berman]

Thank you too!

[@PaulyWollyUTube]

I am running on a Windows 11 environment and when I installed AutogenStudio, I see NO "Playground" tab. Is "Playground" not configured to run on Windows? Every video I see on KZbin has Mac users talking about and using the "Playground" tab... but there is no clear definition of what Windows users should see or how to configure specifically for Windows, to get the "Playground" tab to show.

[@BradBounds]

What did you use and how did you correct the spelling during this video? That is a problem that I have been dealing with for a while now.

[@mikewill7seven]

I feel like the most recent errors everyone is experiencing is as a result of OpenAI putting in some kind of weird limitations, maybe in an effort to force people to use the GPTs. Here is the error I'm receiving. "I apologize for any confusion, but as an AI language model, I don't have the capability to fetch or access external content such as KZbin videos or their transcripts directly. If you have the transcript text, please provide it, and I will be more than happy to assist you in creating a blog post and a tweet thread from it. If you need guidance on how to manually obtain the transcript from a KZbin video, I can provide you with instructions on how to do so. Once you have the transcript, you can paste it here, and we can proceed with crafting your content."

[@nuclearthrone2497]

Try to change the mode name to gpt-4-turbo-preview

[@nuclearthrone2497]

Also weird, you need to add the “code executor” basic agent template to the group with the 2 you created

[@johnt2491]

Matthew I also ran into the same frustrating issue when trying to upgrade AutoGen. Had to use an entirely different dev box to get it working again. 😅 Thanks for the video as always!!

[@friendofai]

Great video! I was able to update a lot easier. I just opened the old version, it told me about the update, and It did it automatically.

[@nsmsms1]

I'm not sure how to write skills for Autogen. Whats the anatomy of a skill? For example, if I want to get information from my email, or scrape my drives, or use an API from a provider, how do I write a skill for that? Can someone help? Matthew, can you create a video for that?

[@shivkumarmallesappa3410]

Do we still need a Model to carry out simple Task. Example Splitting PDF into Multiple Pages. This can be achieved using a simple python code. Do you think Agents are required to run/execute such tasks where the Model or AI does not have any role. Thank you

[@stevenbaert1974]

Thanks! The KZbin content getter works fine on itself but not in the group chat, I get "Since I cannot execute code, I cannot provide the actual transcript or the blog post content. However, if you have access to Python and the necessary libraries, you can use the code snippet provided above to fetch the transcript and follow the subsequent steps to create your blog post." I know latest AutoGen works with Docker by default but then nothing should work right (plotting a stock etc works fine). Any idea?

[@andreasazzi9696]

Please make more tutorials like this in autogen, Matt!!! Thanks a lot!

[@tee_iam78]

Good work, Matt.

[@erojaso]

Amazing tutorial... Hi Matthew. How could we build our own model definition? And how to register it in AutoGen Studio, so we can configure it later. Our business scenarios have multiple layers of security, a simple connection is not enough!

[@franks.6547]

Can an agent be made "librarian" for a local legacy codebase, so that new coding would make use of, and take inspiration from pre-existing code?

[@BalajiPitch]

Keep getting this error on the yt_transcript agent: "I'm sorry for any confusion, but as an AI language model developed by OpenAI, I don't have the capability to execute code or fetch data from external sources", does this work for other using the gpt-4 preview model?

[@draganpetrov9594]

hey matthew, I was wondering how do you save your agents & workflows? everytime I start the UI instance all my work is gone.

[@ioio618]

Hello very informative video...learned a LOT...just a question..is there a way to use the Bard API instead of the OPEN AI one??..should be free at the moment (if I understood well..)

[@NevilHulspas]

What was the use of groupchat_assistant here? It seems like each agent was talking to it, but not sure what its role is compared to user proxy?