AutoGen Update: Complex Tasks and Agents

Adam Fourney discusses the effectiveness of using multiple agents, working together, to complete complex multi-step tasks. He will showcase their capability to outperform previous single-agent solutions on benchmarks like GAIA, utilizing customizable arrangements of agents that collaborate, reason, and utilize tools to achieve complex outcomes.
Microsoft Research Forum, June 4, 2024
See more at aka.ms/ResearchForum-Jun2024
*edited to update a slide

[@MicrosoftResearch]

Discover more about research at Microsoft through our Research Forum Copilot experience: aka.ms/ResearchForumCopilot_YT

[@impolitevegan3179]

Beautiful presentation - concise, authentic, to the point.

[@enespoyraz849]

I, for One, Welcome Our Agent Overlords

[@NowayJose14]

This is great, thanks for all the good work. That being said, let my man take a nap.

[@hawkii1646]

Stop 😂

[@RohanKumar-vx5sb]

Nice! This can be turn key to so many businesses!

[@davorrunje8251]

Great work! 👏👏👏

[@clarkwallace227]

Let’s go Autogen! Keep up the fight.

[@al3030]

Very nice overview. Thanks!

[@kanoptc]

The presenter here is awesome! Use him more!

[@Oliv-B]

Yes a tutorial will be good. Does RAG is recommended for text documents mainly?

[@iseverynametakenwtf1]

So do we have agents that can look at photos now?

[@BurkhardReffeling]

Is there a notebook / example code of what Adam discussed?

[@moisesbessalle]

Do you consider this production ready?

[@austinmw89]

Could AutoGen be used as the backend for generative agents in more complex simulation environments like AI Town? Might it provide useful abstractions for agent memory, perception, reflection and planning?

[@swipekonme]

all that is one way, agents are working on web resources designed for humans. there needs to be a mirror effort where web resources become agent friendly. also this would ultimately result in scalping for instance so we would need circular lottery systems for all qualifying agents within an interaction quantum

[@Malbao14]

Amazing! When can we use it?

[@Neomadra]

It's already open source for a year or so. Just Google autogen

[@RosserNFT]

So cool

[@bubbajones5873]

My favorite part is the crackhead presentation style 👍

[@flor.7797]

we could just put a human in the loop as the operator so basically

[@user-lb2gu7ih5e]

Google "YouSum Live" 00:00:24 Goal: Achieve long-running complex tasks with large models 00:02:05 Multi-agent workflows enhance task decomposition and specialization 00:02:36 Autogen platform facilitates multi-agent collaboration 00:03:16 Four agents achieved top results on Gaia Benchmark 00:03:28 Doubled performance on the hardest set of questions 00:03:49 Agents follow a structured loop involving tasks delegation 00:04:12 Educated guesses crucial for language model speculation 00:05:13 Future plans include introducing agents for self-improvement Google "YouSum Live"