Generalized Multi-Agent Architecture

KTM is an multi-agent architecture where multiple agents collaborate using file system as interface. (e.g. Task Processor agent write output file, which Task Evaluation agent take as an input.) We want to take this idea further to build an architecture where

![[general-agent-archi.png]]

Agents

Agent takes certain input, process it, and create the output. Depending on a situation, input can be one or multiple files, and the same can be said for the output.

Input Certain change in input file system

  • A new file is created
  • Existing file is updated

Process Pre-defined set of actions
Roughly equivalent to existing [[Prompts]]

Output Output file or files with certain specification

Skills

Extract the reusable skills from existing prompts. Load skills as needed for prompt execution.

도메인별 특화 스킬:

  • 📥 컨텐츠 수집: Clippings, Limitless, Photolog 처리
  • 📤 퍼블리싱: Substack, Thread, LinkedIn 채널별 포맷
  • 🗂️ 지식 조직화: Topic 업데이트, Wiki 링크 검증
  • 📝 Obsidian 규칙: 링크 포맷, 폴더 구조, Property 표준

Orchestrator

Orchestrator manage task-level agents to accomplish goals. Orchestrator also interact with user for status reporting and planning improvements.

Task Management Orchestrator monitor the file system and spawn one or more agents to process the task. Orchestrator also check the output of the task and determine whether it’s completed or needs more work.

User Interaction Orchestrator is also interact with user to notify the completion of a task and collect feedback if needed.

Process Improvement Orchestrator also perform routine status report where the issues are diagnosed. Use these results to improve the agent and skill definition over time.

Configurable Repo

Repo will be simplified to contain only

  • Agent Definitions
    • Input/Output location needed to be created
  • Skill Definitions
  • Templates (use existing if needed)
  • Global Agent Rule files

At installation, users can select which agents they like to use, and input/output locations