ํŠธ๋žจ ๋งˆ์Šค์ฝ”ํŠธ
v2.0 โ€” ํŠธ๋žจ Central Model

AI Agent Governance

์ž์œจ ์—์ด์ „ํŠธ์˜ ์šด์˜ ์•„ํ‚คํ…์ฒ˜, ์›Œํฌํ”Œ๋กœ์šฐ ํ”„๋กœํ† ์ฝœ, ๋ฆฌ์†Œ์Šค ๊ณ„์•ฝ์„ ์ •์˜ํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ

๐Ÿ“– ์‹œ์ž‘ํ•˜๊ธฐ โญ GitHub

APEI Protocol

๋ชจ๋“  ์—์ด์ „ํŠธ๊ฐ€ ์ค€์ˆ˜ํ•ด์•ผ ํ•˜๋Š” 4๋‹จ๊ณ„ ์›Œํฌํ”Œ๋กœ์šฐ ํ”„๋ ˆ์ž„์›Œํฌ

Analyze

๐Ÿ” Analyze

์ปจํ…์ŠคํŠธ ๋กœ๋”ฉ ๋ฐ ์š”๊ตฌ์‚ฌํ•ญ ๋ถ„์„

  • ๊ด€๋ จ ํŒŒ์ผ/๋ฌธ์„œ ์‹๋ณ„
  • ๋ชจํ˜ธํ•œ ์  ์งˆ๋ฌธ
  • ์ œ์•ฝ ์กฐ๊ฑด ํŒŒ์•…
Plan

๐Ÿ“‹ Plan

์ž‘์—… ๋ถ„ํ•ด ๋ฐ ์‹คํ–‰ ๊ณ„ํš ์ˆ˜๋ฆฝ

  • 3-10๊ฐœ ํ•˜์œ„ ํƒœ์Šคํฌ ๋ถ„ํ•ด
  • ์˜ํ–ฅ ๋ฒ”์œ„ ๋ถ„์„
  • ์‚ฌ์šฉ์ž/์˜ค์ผ€์ŠคํŠธ๋ ˆ์ดํ„ฐ ์Šน์ธ
Execute

โšก Execute

์›์ž์  ์‹คํ–‰ ๋ฐ ์ƒํƒœ ๋กœ๊น…

  • ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ๋ณ€๊ฒฝ
  • MCP ํ‘œ์ค€ ์ค€์ˆ˜
  • ๋งค ๋‹จ๊ณ„ ์ง„ํ–‰ ์ƒํ™ฉ ์—…๋ฐ์ดํŠธ
Iterate

๐Ÿ”„ Iterate

๊ฒ€์ฆ ๋ฐ ์ž๊ธฐ ์ˆ˜์ •

  • ํ…Œ์ŠคํŠธ/๋ฆฐํŠธ ์‹คํ–‰
  • ์˜ค๋ฅ˜ ์‹œ ์ตœ๋Œ€ 3ํšŒ ์žฌ์‹œ๋„
  • 5ํ„ด ์ดˆ๊ณผ ์‹œ ์—์Šค์ปฌ๋ ˆ์ด์…˜

Deployment Modes

ํ”„๋กœ์ ํŠธ ๋ณต์žก๋„์™€ ์š”๊ตฌ์‚ฌํ•ญ์— ๋”ฐ๋ฅธ ์„ธ ๊ฐ€์ง€ ๋ฐฐํฌ ๋ชจ๋“œ

๐Ÿค–

Single Agent

๋‹จ์ผ LLM์ด ๋ชจ๋“  ์—ญํ•  ์ˆ˜ํ–‰. ์†Œ๊ทœ๋ชจ ์ž‘์—…๊ณผ ๋น ๋ฅธ ํ”„๋กœํ† ํƒ€์ดํ•‘์— ์ ํ•ฉ.

๐Ÿ‘ฅ

CrewAI

์—ญํ•  ๊ธฐ๋ฐ˜ ํ˜‘์—… ์—์ด์ „ํŠธ. ํŒ€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๊ณผ ์ „๋ฌธ ๋ถ„์•ผ ๋ถ„๋ฆฌ์— ์ ํ•ฉ.

๐Ÿ”—

LangGraph

์ƒํƒœ ๊ธฐ๋ฐ˜ ๊ทธ๋ž˜ํ”„ ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜. ๋ณต์žกํ•œ ์›Œํฌํ”Œ๋กœ์šฐ์™€ ์„ธ๋ฐ€ํ•œ ์ œ์–ด.

๐Ÿšƒ ํŠธ๋žจ Central Model (Hybrid)

ํŠธ๋žจ Central Model

Resource Governance

์—์ด์ „ํŠธ์˜ ๋ฌดํ•œ ๋ฃจํ”„ ๋ฐฉ์ง€์™€ ํšจ์œจ์„ฑ์„ ์œ„ํ•œ ์ž์› ๊ด€๋ฆฌ ๋ฐ ๋ฐ์ดํ„ฐ ์ตœ์ ํ™”

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”‚ Constraint โ”‚ Limit โ”‚ On Exceed โ”‚ โ”œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค โ”‚ Iterations per task โ”‚ 5 โ”‚ Escalate โ”‚ โ”‚ Retry on error โ”‚ 3 โ”‚ Escalate โ”‚ โ”‚ Files per change โ”‚ 5 โ”‚ Split task โ”‚ โ”‚ Execution time โ”‚ 10min โ”‚ Checkpoint & pause โ”‚ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

๐Ÿ“Š Data Efficiency: TOON Format

๋Œ€๋Ÿ‰ ๋ฐ์ดํ„ฐ ์ž…์ถœ๋ ฅ ์‹œ TOON(Token-Oriented Object Notation) ํฌ๋งท์„ ์‚ฌ์šฉํ•˜์—ฌ ํ† ํฐ ๋น„์šฉ์„ ์ ˆ๊ฐํ•˜๊ณ  ์ฒ˜๋ฆฌ ์†๋„๋ฅผ ๊ทน๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค.

// TOON (Token-Oriented Object Notation) Example
recipes[2]{id,title,calories}:
1,Kimchi Stew,250
2,Bulgogi,450

๐Ÿง  Context Engineering

ํŒŒํŽธํ™”๋œ ์ •๋ณด ๊ฒ€์ƒ‰(RAG)์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด, ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์กฐ๊ฐ(Chunk)์— ๋ฌธ์„œ ์ „์ฒด์˜ ์ƒํ™ฉ ๋งฅ๋ฝ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์ •ํ™•๋„๋ฅผ ๊ทน๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค.

// Contextual Retrieval Concept
[Original Chunk]: "๋งค์ถœ์ด 3% ์„ฑ์žฅํ•จ"
[Context Added]: "์ด ์ •๋ณด๋Š” 2023๋…„ Q2 ACME Corp์˜ ์žฌ๋ฌด ๋ณด๊ณ ์„œ ๋‚ด์šฉ์ž„. ๋งค์ถœ์ด 3% ์„ฑ์žฅํ•จ"
โ†’ ๊ฒ€์ƒ‰ ์‹คํŒจ์œจ 67% ๊ฐ์†Œ