Post

Let's build multi-agent system

๐ŸŽ‰ Yet another multi-agent framework has arrived!

AgentGym offers a wide range of environments and tasks for broad, real-time, uniform, and concurrent agent exploration!

Following popular releases like CrewAI and Autogen, hereโ€™s another new multi-agent framework. Here are some insights:

  • โ›ณ AgentGym provides diverse environments, tasks, and goals for LLM-based agents with convenient APIs, standardized task specifications, environment settings, and observation/action spaces. This platform supports online evaluation, trajectory sampling, and interactive training.
  • โ›ณ It offers a database with expanded instructions from various environments and tasks, forming a challenging test set for benchmarking, called AgentEval.
  • โ›ณ It includes a uniformly formatted trajectory set, AgentTraj, for training base agents, and AgentTraj-L, an extended version for maximum performance through behavioral cloning.
  • โ›ณThe authors also introduce AgentEvol, a method to explore agent self-evolution beyond previously seen data across tasks and environments. AgentEvol allows agents to evolve by adaptively accessing and utilizing context based on specific tasks, achieving results comparable to state-of-the-art models while consuming fewer resources.

The AgentGym suite, including the platform, dataset, benchmark, checkpoints, and algorithm implementations is available for the community to build and evaluate generally-capable LLM-based agents.

 AgentGym framework

Letโ€™s build a crew of AI Agents

A multi agent system that scrapes the web & automatically writes a blog post for you!

Hereโ€™s what Iโ€™ve used:

I published this work as a Lightning AIโšก๏ธStudio, itโ€™s reads like a blog encapsulating all my code & environment!

Take it for a spin now: https://lightning.ai/lightning-ai/studios/let-s-build-a-crew-of-ai-agents?utm_source=akshay

Translate to Korean

๐ŸŽ‰ ๋˜ ๋‹ค๋ฅธ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ๋„์ฐฉํ–ˆ์Šต๋‹ˆ๋‹ค!

AgentGym์€ ๊ด‘๋ฒ”์œ„ํ•˜๊ณ  ์‹ค์‹œ๊ฐ„์ด๋ฉฐ ๊ท ์ผํ•˜๊ณ  ๋™์‹œ ์—์ด์ „ํŠธ ํƒ์ƒ‰์„ ์œ„ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ํ™˜๊ฒฝ๊ณผ ์ž‘์—…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค!

CrewAI ๋ฐ Autogen๊ณผ ๊ฐ™์€ ์ธ๊ธฐ ์žˆ๋Š” ๋ฆด๋ฆฌ์Šค์— ์ด์–ด ๋˜ ๋‹ค๋ฅธ ์ƒˆ๋กœ์šด ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๋ช‡ ๊ฐ€์ง€ ์ธ์‚ฌ์ดํŠธ์ž…๋‹ˆ๋‹ค.

  • โ›ณ AgentGym์€ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ, ์ž‘์—…, ๋ชฉํ‘œ๋ฅผ ํŽธ๋ฆฌํ•œ API, ํ‘œ์ค€ํ™”๋œ ์ž‘์—… ์‚ฌ์–‘, ํ™˜๊ฒฝ ์„ค์ •, ๊ด€์ฐฐ/ํ–‰๋™ ๊ณต๊ฐ„ ๋“ฑ์œผ๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”Œ๋žซํผ์€ ์˜จ๋ผ์ธ ํ‰๊ฐ€, ๊ถค์  ์ƒ˜ํ”Œ๋ง ๋ฐ ๋Œ€ํ™”ํ˜• ๊ต์œก์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
  • โ›ณ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ ๋ฐ ์ž‘์—…์˜ ํ™•์žฅ๋œ ์ง€์นจ์ด ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค๋ฅผ ์ œ๊ณตํ•˜์—ฌ AgentEval์ด๋ผ๋Š” ๋ฒค์น˜๋งˆํ‚น์„ ์œ„ํ•œ ๋„์ „์ ์ธ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค.
  • โ›ณ ์—ฌ๊ธฐ์—๋Š” ๊ธฐ๋ณธ ์—์ด์ „ํŠธ ํ›ˆ๋ จ์„ ์œ„ํ•œ ๊ท ์ผํ•œ ํ˜•์‹์˜ ๊ถค์  ์„ธํŠธ์ธ AgentTraj์™€ ๋™์ž‘ ๋ณต์ œ๋ฅผ ํ†ตํ•œ ์ตœ๋Œ€ ์„ฑ๋Šฅ์„ ์œ„ํ•œ ํ™•์žฅ ๋ฒ„์ „์ธ AgentTraj-L์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
  • โ›ณ์ €์ž๋Š” ๋˜ํ•œ ์ž‘์—… ๋ฐ ํ™˜๊ฒฝ ์ „๋ฐ˜์— ๊ฑธ์ณ ์ด์ „์— ๋ณธ ๋ฐ์ดํ„ฐ๋ฅผ ๋„˜์–ด ์—์ด์ „ํŠธ ์ž์ฒด ์ง„ํ™”๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ AgentEvol์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. AgentEvol์„ ์‚ฌ์šฉํ•˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ํŠน์ • ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ปจํ…์ŠคํŠธ์— ์ ์‘์ ์œผ๋กœ ์•ก์„ธ์Šคํ•˜๊ณ  ํ™œ์šฉํ•˜์—ฌ ๋ฐœ์ „ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋” ์ ์€ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด์„œ ์ตœ์‹  ๋ชจ๋ธ์— ํ•„์ ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

ํ”Œ๋žซํผ, ๋ฐ์ดํ„ฐ ์„ธํŠธ, ๋ฒค์น˜๋งˆํฌ, ์ฒดํฌํฌ์ธํŠธ ๋ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ตฌํ˜„์„ ํฌํ•จํ•œ AgentGym ์ œํ’ˆ๊ตฐ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ๋นŒ๋“œํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

AI ์š”์› ํฌ๋ฃจ๋ฅผ ๊ตฌ์„ฑํ•ด ๋ด…์‹œ๋‹ค

์›น์„ ๊ธ์–ด๋ชจ์œผ๊ณ  ์ž๋™์œผ๋กœ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ์„ ์ž‘์„ฑํ•˜๋Š” ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ!

๋‚ด๊ฐ€ ์‚ฌ์šฉํ•œ ๊ฒƒ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

  • ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ ๊ตฌ์ถ•์„ ์œ„ํ•œ crewAI
  • ํ˜„์ง€์—์„œ LLM์„ ์ œ๊ณตํ•˜๋Š” Ollama (Llama-3)
  • ๊ฐœ๋ฐœ ๋ฐ ํ˜ธ์ŠคํŒ…์„ ์œ„ํ•œ Lightning AI

๋‚˜๋Š”์ด ์ž‘ํ’ˆ์„ Lightning AI โšก๏ธStudio๋กœ ์ถœํŒํ–ˆ๋Š”๋ฐ, ๋‚ด ๋ชจ๋“  ์ฝ”๋“œ ๋ฐ ํ™˜๊ฒฝ์„ ์บก์Šํ™”ํ•˜๋Š” ๋ธ”๋กœ๊ทธ์ฒ˜๋Ÿผ ์ฝํžŒ๋‹ค!

์ง€๊ธˆ ๋ฐ”๋กœ ์‚ฌ์šฉํ•ด ๋ณด์„ธ์š”: https://lightning.ai/lightning-ai/studios/let-s-build-a-crew-of-ai-agents?utm_source=akshay

This post is licensed under CC BY 4.0 by the author.