New AUTOGEN 3.0 Update | Amazing UI

Tyler AI
8 Sept 202411:36

Summary

TLDRオートジェン・スタジオの最新バージョンに関するこの動画では、視覚的に大きく進化した新機能や改善点について解説しています。特に、ドラッグ&ドロップによるエージェントのワークフローの設定、プロファイラビューによるデバッグ機能の強化、さらには新しいコスト解析やツールの可視化機能が導入され、使いやすさが大幅に向上しています。また、再利用可能なテンプレートやワークフローの簡単なエクスポート・デプロイが可能となり、複雑なマルチエージェントシステムのプロトタイピングと管理が容易になりました。視聴者は、このバージョンに対する期待を高められる内容となっています。

Takeaways

  • 🚀 Autogen Studioの新バージョンは大幅に改善され、前バージョンよりも多くの新機能が追加されています。
  • 🖥️ ドラッグ&ドロップUIを使用して、エージェントワークフローを簡単に指定・評価できるようになりました。
  • 📊 新機能として、メトリクスの可視化やメッセージのプロファイリングが導入され、エージェントの動作をより詳細に確認できるようになりました。
  • 📂 Autogen Studioでは、複数のエージェント間でスキルやツールを簡単に割り当て可能になっています。
  • 🛠️ デバッグがより容易になり、エージェントの失敗原因やワークフローコストが明確に表示されます。
  • 🗂️ 再利用可能なエージェントコンポーネントのギャラリーがあり、テンプレートの再利用と共有が可能です。
  • 👨‍💻 コード不要で、Pythonアプリケーションにワークフローをエクスポート・デプロイできる新機能が追加されています。
  • 🎯 Playgroundビューを使用してタスク実行やデバッグができ、ワークフローを迅速にプロトタイプ化できます。
  • 🤖 Autogen Studioは、マルチエージェントシステムのデザインと最適化に向けた新しいツールを提供しています。
  • 📈 プロファイラビューで、エージェントの動作と成果物の生成がすべて可視化され、複雑なシステムでも効率的に管理できます。

Q & A

  • Autogen Studioとは何ですか?

    -Autogen Studioは、マルチエージェントワークフローのプロトタイピング、デバッグ、および評価を迅速に行うためのノーコード開発ツールです。直感的なドラッグ&ドロップUIを提供し、エージェントのワークフローを簡単に設計できます。

  • Autogen Studioの最新バージョンで何が改善されましたか?

    -最新バージョンでは、ドラッグ&ドロップによるUIの大幅な改善、プロファイリング機能の追加、エージェント間のメッセージやアクションの可視化、コストの追跡などが強化されました。

  • 新バージョンのAutogen Studioでどのような機能が追加されましたか?

    -エージェント間のメッセージやアクションの可視化、タスクの進行状況の監視、ワークフローのコスト計算、ツールの成功・失敗を追跡するプロファイリング機能などが追加されました。

  • エージェントワークフローを作成する際にどのようなモデルやツールを使用できますか?

    -GPT-4 Turboなどのモデルを選択して使用することができ、コンテンツ生成、画像生成、ウェブ検索などのスキルをエージェントに追加することが可能です。

  • Autogen Studioの新しいプロファイリング機能とは何ですか?

    -プロファイリング機能は、エージェントのメッセージ、アクション、コスト、成功したツールや失敗したツールを可視化し、ワークフローのデバッグや最適化をサポートします。

  • Autogen Studioはどのようにワークフローのエクスポートをサポートしていますか?

    -Autogen Studioでは、ワークフローをJSON形式でエクスポートでき、Pythonアプリケーションで実行したり、Dockerコンテナとしてデプロイしたりすることが可能です。

  • Autogen Studioの最新バージョンでの課題は何ですか?

    -マルチエージェントワークフローの設計は複雑であり、特にエージェントの失敗や問題点を特定するのが困難ですが、新しいプロファイリングツールでこれを解決しようとしています。

  • Autogen Studioで作成されたワークフローのデバッグ方法はどのように改善されましたか?

    -デバッグ機能が強化され、エージェントが実行するタスクの進行状況や生成されたファイル(画像、コード、ドキュメントなど)をリアルタイムで観察できるようになりました。

  • Autogen Studioの『Playground View』とは何ですか?

    -Playground Viewは、タスクの実行やワークフローのデバッグを行うためのインターフェースで、エージェントの動作や生成物を視覚的に確認できます。

  • Autogen Studioのエージェントはどのような用途に使用できますか?

    -エージェントは、ユーザープロキシ、コンテンツ生成、画像生成、QAエージェントなど様々な役割を持つことができ、グループチャットやドキュメント作成などのタスクに応用可能です。

Outlines

00:00

🛠️ オートジェンスタジオの大幅な改善

オートジェンスタジオの新バージョンについての紹介。このツールはマルチエージェントワークフローのプロトタイプ作成、デバッグ、評価を簡単に行えるノーコード開発ツールです。新バージョンはドラッグ&ドロップのインターフェースを採用し、直感的な操作性が大幅に向上しています。オートジェンスタジオの目標は、より複雑なシステムを管理しやすくすることです。エージェントを使ったプロセスの効率化が期待されており、プロファイリング機能や可視化が追加され、ワークフロー全体のコストやエージェント間のやりとりが一目でわかるようになっています。

05:02

🎛️ プレイグラウンドビューとワークフローのデバッグ

プレイグラウンドビューは、タスク実行やワークフローのデバッグに使われる新機能で、エージェントの行動を簡単に追跡できます。このビューでは、セッションを作成し、ワークフローを割り当て、タスクを実行できます。ワークフローは、単発タスクやマルチターンのタスクとして動作します。デバッグをサポートするために、タスクの進行状況やエージェントのメッセージがリアルタイムで表示され、生成された成果物が確認できます。また、新しいインターフェースが導入され、ユーザーはより多くの情報を容易に取得できるようになっています。

10:03

📘 新しい機能とエクスポートの改善

オートジェンスタジオでは、ワークフローをJSONファイルとしてエクスポートでき、Pythonアプリケーションに直接組み込むことが可能になりました。また、エージェント間のやり取りの可視化やコスト計算の詳細な表示など、全体的な操作性が大幅に改善されています。デバッグやシステムの挙動を理解するためのツールが強化され、複数エージェントによるワークフローの作成が簡単になりました。このバージョンでは、エージェントを迅速にプロトタイピングできるだけでなく、将来の研究やデザインパターンの向上に向けた方向性も示されています。

Mindmap

Keywords

💡Autogen Studio

Autogen Studioは、マルチエージェントワークフローの迅速なプロトタイピング、デバッグ、および評価をサポートするノーコード開発ツールです。このツールは、エージェントのワークフローを直感的なドラッグアンドドロップUIで構築できる点が強調されています。動画内では、新バージョンが大幅に改善され、特に視覚的なインターフェースやエージェント間の相互作用がわかりやすくなったことが示されています。

💡マルチエージェントワークフロー

複数のエージェントが協力してタスクを実行するシステムを指します。Autogen Studioでは、このマルチエージェントワークフローを容易に設計し、デバッグできる点が強調されています。ビデオでは、各エージェントが独自の役割を果たし、エージェント間の相互作用を視覚的に確認できる機能が紹介されています。

💡ドラッグアンドドロップUI

Autogen Studioの新しい機能として、エージェントやスキルを簡単にドラッグアンドドロップで配置するUIが導入されました。この操作により、ユーザーはプログラミング知識がなくてもワークフローを直感的に構築できます。ビデオでは、エージェントやツールの追加がどれだけ簡単になったかが強調されています。

💡プロファイリング機能

プロファイリング機能は、各エージェントがどのようにタスクを実行したか、どのツールを使用したか、コストはどれくらいかかったかなどの詳細な情報を提供します。特に、ワークフローのデバッグ時に重要な役割を果たし、失敗した箇所やその原因を特定するのに役立ちます。

💡タスクデバッグ

タスクデバッグは、エージェントがタスクを正しく実行できるかを確認し、問題が発生した場合にその原因を探るためのプロセスです。Autogen Studioでは、視覚的なフィードバックを通じて、エージェントの動作を簡単に追跡できる機能が追加されました。

💡エージェントテンプレート

再利用可能なエージェントテンプレートは、開発者が同様のタスクを迅速に設定できるように設計されています。ビデオでは、テンプレートがギャラリーで提供され、ユーザーがそれを共有したり、特定のワークフローに適用できることが示されています。

💡ノーコード開発

プログラミングの知識がなくても、視覚的な操作だけでワークフローを構築できる開発方法を指します。Autogen Studioでは、ユーザーはコードを一切記述せずに複雑なマルチエージェントワークフローを作成できる点が強調されています。

💡エージェント間のメッセージ可視化

エージェントが相互にどのようなメッセージを送受信しているかを視覚的に表示する機能です。この機能により、タスクの進行状況や問題点を簡単に把握できるようになります。ビデオでは、メッセージの数やツールの呼び出しの成否が表示される様子が示されています。

💡ワークフローのエクスポート

Autogen Studioでは、作成したワークフローをJSON形式でエクスポートでき、それをPythonアプリケーションやAPIエンドポイントで使用できるようになっています。これにより、開発者は柔軟にワークフローを他のシステムに統合できます。

💡GitHubイシューの可視化

GitHubのリポジトリにおけるユーザーフィードバックやイシューが視覚化され、問題点や改善点が一目でわかるようになります。Autogen Studio 3.0では、この機能を活用してユーザーからのフィードバックを整理し、次期バージョンの改善に役立てています。

Highlights

Huge improvement from the previous version of Autogen Studio.

Introduction of new features with a more intuitive drag-and-drop UI.

Built off the Autogen framework developed over the past year.

No-code developer tool for rapidly prototyping, debugging, and evaluating multi-agent workflows.

Provides reusable agent components and new profiling capabilities.

Significant visual changes to the interface with added tools for debugging.

Allows drag-and-drop functionality to assign tasks and skills to agents.

New feature that tracks costs, tools executed by agents, and workflow outputs.

Debugging multi-agent workflows with improved observability.

Supports long-term memory and vector databases integration.

Playground view for task execution and workflow debugging.

Supports reusable templates for agents like assistant agents and group chat agents.

New gallery view for sharing and reusing templates.

Support for exporting workflows as JSON config files.

Workflows can be deployed in Python applications or wrapped in Docker containers.

Transcripts

play00:00

I have to say this looks absolutely

play00:02

amazing this is such a huge improvement

play00:04

from the previous version of Auden

play00:06

Studio it seems that it's been a little

play00:07

while since we really heard something

play00:09

about autogen Studio or at least

play00:10

anything significant until now I'm

play00:13

really excited to go over this paper

play00:14

with you and see what all there is with

play00:16

the new version so let's get into it

play00:18

okay so this is the new paper and as you

play00:20

can see just from this screenshot that

play00:22

is going to look so much different than

play00:24

what it currently does mainly because

play00:26

there are new features but let's just go

play00:28

over if you're new to aen Studio and

play00:30

what it is let's just go over what this

play00:32

paper is saying and we'll get to these

play00:33

features the idea of autogen studio is

play00:36

that it's built off the autogen

play00:38

framework which has developed over the

play00:39

past year into something wonderful

play00:42

however with that comes complexity right

play00:44

the more you're adding onto the

play00:45

framework to do more the more complex it

play00:48

can get and this is where autogen Studio

play00:50

comes into play so to address this

play00:53

situation they present aen Studio A no

play00:55

code developer tool for rapidly

play00:57

prototyping debugging and evalu evalua

play01:00

these multi-agent workflows it provides

play01:02

an intuitive drag and drop UI for agent

play01:05

workflow specification interacting the

play01:07

evalu interactive valuation and

play01:09

debugging of workflows and a gallery of

play01:11

reable agent components and if you see

play01:14

with a screenshot right this is nothing

play01:16

at all what autogen studio used to look

play01:18

like this is now a drag and drop see

play01:21

there's an agent a there's an agent B

play01:23

look this is the um the agent a is a

play01:26

user proxy and it is executing code B is

play01:29

a book generation group chat manager and

play01:31

inside of here there are three different

play01:33

agents and it looks like they're just

play01:34

dragging and dropping the model so right

play01:37

here is GPT 4 Turbo they're just

play01:38

dragging dropping the models and adding

play01:40

skills right so this one's getting a

play01:43

Content agent is getting web search and

play01:44

image agent is getting an image

play01:46

generator skill or tool and I have to

play01:49

say that is an this this screenshot

play01:51

right here is an amazing Improvement of

play01:54

what it used to look like now that I see

play01:56

this I'm loving this but what's

play01:58

interesting is this introduction of

play02:00

profiling capabilities with

play02:02

visualizations of messages and actions

play02:05

and well this metrics right here metrics

play02:06

is the main thing here is it looks like

play02:09

they're they're going to have a UI to

play02:11

show you um you know what the costs are

play02:14

what tools were executed by which agents

play02:17

what are the output for each of the

play02:19

agent and so forth so if something if

play02:22

something failed you can now see where

play02:24

it was and then also how much your

play02:26

workflow cost you and it is hard to

play02:28

debug some of these mod multi- agent

play02:30

workflow especially the more complex

play02:32

they get right if something fails yeah I

play02:34

mean you can have you can have logging

play02:36

and print statements and what or however

play02:37

you do it but still to see exactly how

play02:41

the agent that specific agent failed and

play02:43

why it failed it can be difficult so

play02:46

looks like here they have a related work

play02:47

section where they talk about uh agentic

play02:50

implementation such as react which is

play02:53

the reason and acting for llms they have

play02:56

L chain for Harrison Chase uh and so

play03:00

forth this is just kind of um saying

play03:04

about what some people have done and

play03:05

some limitations to have better

play03:08

architectures for multi-agents all right

play03:10

and I'm just going to go over these

play03:11

Concepts really quick just in case you

play03:13

know some of you a lot of you may

play03:14

already know these but these are like

play03:16

going to be the main components of

play03:17

autogen Studio which is why they're

play03:19

highlighting here right it's this whole

play03:20

drag and drop experience now so with the

play03:23

with the model you know this is you're

play03:25

going to be able to pick the model to

play03:26

generate whatever it is image text uh

play03:29

skills SL tools you know you're going to

play03:31

I think you can it looks like you can

play03:32

still if you want to code your own tool

play03:35

and then you can probably drag and drop

play03:37

that onto whichever agent and it looks

play03:39

like they may be implementing memory so

play03:41

shortterm you know this is the normal or

play03:44

long-term right so they can we can have

play03:46

Vector databases which would be nice to

play03:48

have um with autogen Studio I will'll

play03:50

look further in the research paper see

play03:52

if they're going to implement that but

play03:53

that would be really nice to have and of

play03:55

course the agent and then you can looks

play03:57

like you can combine um like workflows

play04:01

together to create these groups of

play04:02

Agents or like a group chat or you can

play04:04

have a single agent or whatever it is

play04:06

you know and scrolling down a little bit

play04:07

here what they're kind of just getting

play04:09

at is that you know there are

play04:11

limitations um whenever you have all of

play04:13

these together that like you know it can

play04:15

be hard to basically have reusable

play04:18

templates or be able to bootstrap some

play04:21

workflow you know it can be hard so they

play04:24

try to get rid of that just by providing

play04:26

this visual interface and they're going

play04:27

to and they're going to kind of hone in

play04:29

on the design goals which is to have

play04:31

rapid prototyping developer tooling so

play04:33

these tools um you know we can help

play04:37

understand what's happening with agent

play04:38

behaviors and this is going to help us

play04:40

you know facilitate the Improvement of

play04:42

these systems by understanding quicker

play04:45

and easier what's happening and they

play04:47

have reusable templates now this they

play04:49

had a gallery section right that present

play04:51

a gallery of reusable sharable templates

play04:54

it could be just still in the same

play04:55

Gallery section but you know that was a

play04:57

little different so hopefully it's

play04:59

improved with um with this next version

play05:01

of autogen studio so they have a

play05:03

playground view which is going to be

play05:05

used for tax task execution workflow

play05:08

debugging and options to export and

play05:10

deploy the gallery view fac facilitates

play05:13

the ReUse and sharing of the templates

play05:16

which is what I just talked about now

play05:17

this kind of confirms it and then here

play05:20

they're you know they're just going over

play05:22

some templates they have for um for

play05:25

agents like assistant agent a group chat

play05:27

uh agent and so forth right this is

play05:30

that's all pretty standard stuff with

play05:31

autogen saying workflows can be tested

play05:33

in the build view which you know we

play05:36

could we did already um in the previous

play05:38

version I believe or was made the

play05:39

playground view but more systematically

play05:42

explored within the playground view okay

play05:44

so the playground view allows users to

play05:46

create sessions which again we were able

play05:49

to do in the previous version attach

play05:50

workflows to the session and then run

play05:53

tasks so either single shot or

play05:54

multi-turn basically group chats or just

play05:57

a onetoone um or just an Interactive in

play05:59

with one agent aen Studio provides two

play06:02

features to support debugging first it

play06:04

provides an observe view whereas task

play06:06

progress messages and actions performed

play06:08

by agents are streamed to the interface

play06:09

and all generated artifacts are

play06:12

displayed so files uh such as images

play06:15

code and documents and now let's go

play06:17

ahead and look at the new interface and

play06:20

here all they're really saying is the

play06:21

backend API the frontend web UI and kind

play06:24

of what it looks like but look at just

play06:27

look at how this looks this looks abs

play06:29

absolutely amazing this is a huge

play06:31

upgrade from the way it used to look and

play06:34

the amount of information that you can

play06:35

gather from this so it looks like

play06:37

they're in the playground session right

play06:39

now and so they asked you to create a

play06:41

children's uh PDF book with four pages

play06:44

each describing the weather in Seattle

play06:46

so then the it says the agents have

play06:49

completed the task right so this again

play06:51

this was uh part of the group chat

play06:53

manager the all these agents so it's

play06:55

saying the agents have completed the

play06:57

task and then the children's PDF book

play06:59

titled weather in Seattle uh has been

play07:02

created with descriptions and everything

play07:04

the book should now be available as

play07:06

Seattle weather children book.pdf on

play07:09

your system you can now View and ensure

play07:12

that it meets your expectations if

play07:13

everything looks good that completes our

play07:15

task if you need any further assistance

play07:17

or modifications please let me know so

play07:19

it gave a total of seven files so I

play07:21

guess just uh like it created the

play07:24

children's book compiled with seven

play07:25

images and and on the right side here is

play07:27

talking about the profiler which which

play07:29

now it says the group chat manager right

play07:30

these the amounts of tokens that were

play07:32

generated it cost 15 cents the uh then

play07:35

it breaks it down you know so the

play07:37

content this is probably that the agent

play07:40

that was creating the images the user

play07:42

proxy or no I'm sorry here's the image

play07:44

generator uh you know this is only 1

play07:46

cent and the content is 12 cents so I

play07:48

can't imagine the image generator only

play07:50

being one penny um but then they have a

play07:53

QA um agent so this is it this is

play07:56

amazing right the way this looks and

play07:58

then um how many the total messages

play08:00

between all of the agents so it looks

play08:02

like maybe they were uh between the user

play08:05

proxy and the group chat there were

play08:06

probably you know here's about 17 and

play08:09

then the user proxy probably had like

play08:10

six so you know a fair amount of you

play08:14

know it's been a fair amount of messages

play08:16

sent from everybody and then tool and

play08:18

then here's all the tool calls right it

play08:20

says which ones were successful and

play08:22

which ones failed so you know it looks

play08:24

like uh it some of these failed and the

play08:27

thing about this is it's telling you and

play08:29

giving you the metrics for what has

play08:31

happened with this right that's the

play08:32

amazing part about this and now you can

play08:35

deploy workflows auten Studio does

play08:37

enable users to export workflows as a

play08:40

Json config file which they kind of did

play08:42

have before but the what they what

play08:44

you're able to do now is an exported

play08:47

workflow uh can be put into any python

play08:49

application right so if you just have

play08:51

some python code or you know how to do

play08:53

that then you can execute this uh as an

play08:56

API endpoint using the autogen studio

play08:59

CLI or wrapped in a Docker container

play09:03

right so they you know you just you

play09:05

don't have to actually um pip install

play09:07

autogen you can just do from autogen

play09:10

Studio import workflow manager you can

play09:12

give it the Json file that you exported

play09:14

and then it can run it right and then um

play09:17

you know this is basically the workflow

play09:18

that you had created so you can say what

play09:20

is the height of the Eiffel Tower for

play09:22

instance this is probably just a simple

play09:24

onetoone agent interaction but I mean

play09:27

this is such an improvement of what you

play09:28

were able to do and then now finally

play09:31

we're going to come down to the usage

play09:33

and evaluation and what it looks like

play09:36

here is the this is the autogen studio

play09:39

uh GitHub issue visualization right so

play09:42

they kind of grouped uh everything

play09:43

together and you know said the plot of

play09:45

GitHub issues from the autogen studio

play09:47

repo user feedback range from support

play09:50

with workflow uh authoring tools to

play09:53

General installation so they kind of

play09:54

grouped everything together and probably

play09:56

tried to figure out how can we solve the

play09:57

problem for everybody that was kind of

play09:59

that we were kind of having with autogen

play10:01

Studio because at the end of the day it

play10:02

was kind of I don't want to say bare it

play10:05

did get you it allowed you to prototype

play10:07

um agents really quickly but even like

play10:10

after you did that right there wasn't

play10:13

much it felt like there wasn't much more

play10:15

you can do with it like something was

play10:17

kind of missing and it looks like

play10:19

they're trying to fix all of that with

play10:20

this new version of autogen and then

play10:23

we're they're talking about design

play10:24

patterns and research directions um so

play10:27

you know we kind of already talked about

play10:29

the Define and compose workflow this is

play10:31

allows users to author workflows by

play10:33

basically dragging and dropping

play10:35

everything together to create the

play10:37

multiagent workflow and then the

play10:38

debugging and sensemaking tools um this

play10:41

is what we've also talked about like the

play10:43

profiler views all the metrics these

play10:46

help you debug interpret and uh

play10:48

understand the behavior and output of

play10:50

your systems using the no code method

play10:52

right again because this is no code you

play10:54

don't have the typical logging methods

play10:56

which aren't easy with multi-agents so

play10:58

they came up with something so that we

play11:00

could have we could see more of what's

play11:02

Happening under the hood you know and

play11:04

then they have export and deployment

play11:06

there's uh collaboration and sharing you

play11:08

know and how do we understand uh the

play11:10

multi-agent system designs you know and

play11:13

then optimizing the multi-agent systems

play11:15

you know so this is kind of the the

play11:17

questions that they're asking that they

play11:18

want to improve upon in the future well

play11:20

this is autogen 3.0 and I would actually

play11:22

say that because this is such an

play11:24

improvement from the previous version I

play11:26

am really excited for this I hope you

play11:28

are excited too and in the meantime I

play11:30

have a couple courses you can take

play11:32

related to Ai and one of them is autogen

play11:34

thank you for watching I'll see you next

play11:36

video

Rate This

5.0 / 5 (0 votes)

関連タグ
AIツールプロトタイピングドラッグ&ドロップエージェント管理デバッグビジュアルUIワークフローメトリクスノーコードコスト評価