Prompt Engineering for Agent Orchestration

🎯 Overview

Junis uses prompt-based orchestration where you control agent behavior entirely through system prompts (instruction templates). No code changes needed - just configure your prompts in the Admin UI. This guide provides production-tested patterns that work directly in your Junis system.

🧠 Core Concepts

How Junis Routing Works

User Message → Orchestrator (analyzes) → Routes to Sub-Agent → Returns Response

Key Point: Orchestrator decides routing based on keywords in its system prompt.

Where to Configure

Admin UI → Agent Management → Select Agent
Edit Agent → Instruction Template field
Paste your prompt → Save
Done! Agent automatically reloads

🔀 A. Sub-Agent Routing

How It Works

Mention sub-agent names in your Orchestrator’s instruction template, and the LLM will route requests to them based on keywords.

✅ Working Example (From Production)

This is the actual default Orchestrator prompt in Junis:

당신은 Junis의 메인 오케스트레이터입니다.

사용자 요청을 분석하고 적절하게 응답하세요:

1. 이메일 관련 요청 → EmailWorkflow 에이전트로 전송
   - "이메일 작성해줘", "메일 답장", "이메일 템플릿" 등
   - 예: "이메일 작성을 도와드리겠습니다. EmailWorkflow 에이전트로 연결합니다."

2. 일반 대화 → 직접 응답
   - 인사, 일상 대화, 간단한 질문 등
   - 친근하고 전문적인 톤으로 직접 응답하세요.

📋 Pattern Breakdown

Structure:

[Condition Description] → [Sub-Agent Name] 에이전트로 전송
   - [Keyword 1], [Keyword 2], ...
   - 예: [Example response]

Why It Works:

✅ Explicit agent names (EmailWorkflow, etc.)
✅ Clear keywords for pattern matching
✅ Examples guide the LLM’s response style

🛠️ How to Set Up

Step 1: Create Sub-Agents

Admin UI → Agent Management → Create Agent
Create your sub-agents:
- EmailWorkflow (LLM_AGENT)
- DataWorkflow (SEQUENTIAL_AGENT)

Step 2: Link to Orchestrator

Select your Orchestrator agent
Relationships tab → Add Relationship
Add each sub-agent with order index (1, 2, 3…)

Step 3: Configure Orchestrator Prompt

Edit Orchestrator → Instruction Template
Paste routing logic (see example above)
Save → Cache automatically refreshes

✨ Advanced Routing Patterns

Multi-Language Routing

You are a multilingual orchestrator.

Route based on request type:

1. Code requests → CodeAssistant agent
   - "write code", "debug", "refactor", "implement"
   - Example: "I'll connect you to CodeAssistant for coding help."

2. Data analysis → DataAnalyst agent
   - "analyze data", "create chart", "statistics", "trend"
   - Example: "Routing to DataAnalyst for data insights."

3. General chat → Respond directly
   - Greetings, small talk, questions
   - Friendly and professional tone.

Priority-Based Routing

우선순위 기반 라우팅:

1. 긴급 요청 (최우선) → UrgentSupport 에이전트
   - "긴급", "빠르게", "지금 당장", "urgent" 등

2. 기술 문제 → TechSupport 에이전트
   - "에러", "버그", "오류", "작동하지 않음" 등

3. 일반 문의 → GeneralSupport 에이전트
   - 그 외 모든 요청

🛠️ B. Tool Calls & MCP Integration

How It Works

Mention tool/MCP names in your agent’s instruction, and the LLM will call them when needed.

✅ Working Example (Tool Call)

당신은 브랜드 데이터 검색 전문가입니다.

사용 가능한 도구:
1. search_brand_in_systems - 여러 시스템에서 브랜드 정보 검색
2. get_bbpi_data - BBPI 데이터 조회
3. get_cr_segments - CR 세그먼트 데이터 조회

작업 절차:
1. 사용자가 브랜드 이름을 언급하면 search_brand_in_systems 도구 사용
2. 브랜드가 확인되면 get_bbpi_data로 상세 정보 조회
3. 결과를 친절하게 요약하여 제공

예시:
- 사용자: "아모레퍼시픽 정보 알려줘"
- 행동: search_brand_in_systems("아모레퍼시픽") 실행

📋 Pattern Breakdown

Structure:

사용 가능한 도구:
1. [tool_name] - [tool_description]
2. [tool_name] - [tool_description]

작업 절차:
1. [Condition] → [tool_name] 사용
2. [Next step] → [tool_name] 사용

🔧 How to Configure Tools

Step 1: Register Tool in Admin UI

Admin UI → Tools Management
Check available tools (system admin only)
Note the exact function name

Step 2: Add to Agent Config

Agent Management → Select agent → Edit

Tools field → Add tool names:

["search_brand_in_systems", "get_bbpi_data"]

Save

Step 3: Update Instruction Template

Include tool names and usage patterns in the instruction (see example above).

✨ MCP Integration Example

You are a development assistant with GitHub integration.

Available MCP platforms:
1. GitHub - Create issues, manage repositories, and track PRs

Tasks you can perform:
- "create an issue for X" → Use GitHub MCP to create issue
- "list open PRs" → Use GitHub MCP to retrieve pull requests
- "get repository info" → Use GitHub MCP to fetch repo details

Example workflow:
1. User: "Create an issue for bug in authentication"
2. Action: Call GitHub MCP create_issue function
3. Response: "Issue #123 created successfully in repository"

🔧 How to Configure MCP

Step 1: Enable MCP Platform

Admin UI → MCP Platforms
Check available platforms (GitHub, Firecrawl, etc.)
Configure credentials (API keys, OAuth tokens)

Step 2: Link MCP to Agent

Agent Management → Select agent → MCP tab
Add MCP Platform → Select platform
Priority Order: Lower number = higher priority

Step 3: Update Instruction Template

Mention MCP platform capabilities in your prompt (see example above).

📚 C. RAG Datastore Queries

How It Works

Mention data types or topics in your instruction, and the LLM will query your RAG Datastore when needed.

✅ Working Example

당신은 회사 문서 검색 전문가입니다.

사용 가능한 지식 베이스:
- 회사 정책 및 규정 문서
- 제품 매뉴얼 및 사양서
- 고객 FAQ 및 문의 이력

작업 방식:
1. 사용자 질문 분석
2. 관련 문서 자동 검색 (RAG)
3. 검색 결과 기반 답변 생성

예시:
- 사용자: "휴가 정책 알려줘"
- 행동: 회사 정책 문서에서 휴가 관련 내용 검색 → 답변 생성

📋 Pattern Breakdown

Structure:

사용 가능한 지식 베이스:
- [Knowledge domain 1]
- [Knowledge domain 2]

작업 방식:
1. 질문 분석
2. 자동 검색 (RAG)
3. 답변 생성

🗂️ How to Set Up RAG

Step 1: Create DataStore

Admin UI → Knowledge Base → Create DataStore
Name: company-policies (use lowercase with hyphens)
Upload documents: PDFs, DOCX, TXT files

Step 2: Add RAG Tool to Agent

Agent Management → Select agent → Edit
RAG Tools field → Add:
["rag_company-policies"]
Format: rag_ prefix + DataStore name
Save

Step 3: Update Instruction Template

Mention knowledge domains in your prompt (see example above).

⚠️ Important Notes

Model Requirement: RAG only works with Gemini models (e.g., gemini/gemini-1.5-flash)
DataStore Location: Must be in global location (not regional)
Naming: Use lowercase, hyphens only (e.g., rag_brand-knowledge)

🔄 D. Output Format & Key Passing

How It Works

Define output format in parent agent, then reference output key in child agent to receive structured data.

✅ Working Example (Parent Agent)

당신은 데이터 수집 코디네이터입니다.

작업:
1. 여러 소스에서 브랜드 데이터 수집
2. 결과를 JSON 형식으로 출력

출력 형식 (중요):
{
  "brand_name": "브랜드명",
  "bbpi_data": { BBPI 상세 정보 },
  "cr_data": { CR 세그먼트 정보 },
  "salesforce_data": { Salesforce CRM 정보 }
}

이 출력은 ReportCompiler 에이전트가 사용합니다.

DB Configuration:

Output Key: collected_brand_data

✅ Working Example (Child Agent)

당신은 데이터 분석 보고서 컴파일러입니다.

입력 데이터:
- 이전 에이전트의 collected_data 출력을 받습니다.

작업:
1. collected_data에서 핵심 정보 추출
2. 분석 보고서 형식으로 재구성
3. 마크다운 또는 HTML로 최종 보고서 생성

출력 형식:
# [데이터 이름] 분석 보고서

## 1. 데이터 분석
[collected_data.analysis 내용]

## 2. 세그먼트 정보
[collected_data.segments 내용]

## 3. CRM 정보
[collected_data.crm_data 내용]

DB Configuration:

Include Contents: collected_brand_data

📋 Pattern Breakdown

Parent Agent:

출력 형식 (중요):
{
  "key1": "value description",
  "key2": { nested object },
  ...
}

이 출력은 [ChildAgent] 에이전트가 사용합니다.

Child Agent:

입력 데이터:
- 이전 에이전트의 [output_key] 출력을 받습니다.

작업:
1. [output_key]에서 [data] 추출
2. [Process]

🔧 How to Configure

Step 1: Parent Agent Setup

Agent Management → Select parent agent → Edit
Instruction Template: Define output format
Output Key field: collected_brand_data
Output Schema (optional): Define JSON schema
Save

Step 2: Child Agent Setup

Agent Management → Select child agent → Edit
Instruction Template: Mention input data source
Include Contents field: collected_brand_data
Save

Step 3: Link Agents

Create Sequential Agent (e.g., DataWorkflow)
Add relationships:
- Order 1: DataCollector (parent)
- Order 2: ReportCompiler (child)

✨ Advanced: Schema Validation

Define output schema for strict validation:

{
  "type": "object",
  "properties": {
    "brand_name": { "type": "string" },
    "bbpi_data": { "type": "object" },
    "cr_data": { "type": "object" }
  },
  "required": ["brand_name", "bbpi_data"]
}

Add this to Output Schema field in Admin UI.

🎭 E. Flexible Application Patterns

Conditional Logic

당신은 스마트 라우터입니다.

조건부 라우팅:

IF 사용자 요청에 "긴급" 또는 "urgent" 포함:
  → UrgentHandler 에이전트로 즉시 전달

ELSE IF 사용자가 파일 첨부:
  → FileProcessor 에이전트로 전달

ELSE IF 질문이 회사 정책 관련:
  → 지식 베이스 검색 수행

ELSE:
  → 일반 응답 제공

Multi-Step Workflows

당신은 워크플로우 관리자입니다.

단계별 작업 프로세스:

1단계: 데이터 검증
   - 사용자 입력 확인
   - 필수 필드 체크
   - 유효성 검사

2단계: 외부 시스템 조회
   - Salesforce에서 고객 정보 조회 (MCP)
   - DataLab에서 거래 이력 조회 (Tool)

3단계: 분석 및 보고서 생성
   - 수집된 데이터 통합
   - 인사이트 도출
   - PDF 보고서 생성 (Tool: claude_pdf_processor)

4단계: 결과 전달
   - GitHub 이슈 생성 또는 커밋
   - 사용자에게 완료 알림

Dynamic Tool Selection

당신은 유연한 도구 사용자입니다.

상황에 맞는 도구 선택:

데이터 소스 우선순위:
1. 먼저 내부 데이터베이스 검색 (get_internal_data 툴)
2. 없으면 Salesforce 조회 (Salesforce MCP)
3. 그래도 없으면 웹 검색 수행 (Web Search 활성화 필요)

항상 가장 정확한 데이터 소스를 먼저 시도하세요.

Error Handling

당신은 안정적인 에이전트입니다.

에러 처리:

IF 도구 호출 실패:
  → 사용자에게 친절히 설명
  → 대체 방법 제안

IF 데이터 없음:
  → "해당 정보를 찾을 수 없습니다" 명확히 전달
  → 추가 정보 요청 또는 다른 접근 방법 제시

IF 외부 시스템 오류:
  → 재시도 1회
  → 실패 시 수동 대응 안내

✅ Best Practices

1. Be Explicit

❌ Bad: “Handle email requests” ✅ Good: “이메일 작성, 답장, 전달 요청 → EmailWorkflow 에이전트로 전송”

2. Provide Examples

Include example user inputs and agent responses in your prompts.

3. Use Clear Formatting

Structure your prompts with:

Numbered lists
Clear headers
Example blocks

4. Test Incrementally

Test Orchestrator routing first
Add one sub-agent at a time
Verify each tool/MCP call individually
Test complete workflow last

5. Monitor & Iterate

Check Usage Logs in Admin UI
Review agent routing events
Refine prompts based on real usage

🐛 Troubleshooting

Agent Not Routing

Problem: Orchestrator responds directly instead of routing. Solution:

✅ Check sub-agent name matches exactly in prompt
✅ Verify agent relationship exists in Admin UI
✅ Ensure sub-agent is active (not disabled)

Tool Not Called

Problem: Agent doesn’t call the tool. Solution:

✅ Verify tool is added to Tools field in agent config
✅ Mention tool name and usage in instruction
✅ Check tool is registered in system (Tools Management)

MCP Call Fails

Problem: MCP platform error or timeout. Solution:

✅ Verify MCP credentials in MCP Platforms settings
✅ Check MCP platform is linked to agent (MCP tab)
✅ Test MCP connection manually in Admin UI

RAG Returns No Results

Problem: RAG query returns empty or irrelevant results. Solution:

✅ Verify DataStore has indexed documents
✅ Check agent uses Gemini model (RAG only works with Gemini)
✅ Ensure RAG tool name format: rag_[datastore-name]
✅ Improve query phrasing in prompt

Output Key Not Passed

Problem: Child agent doesn’t receive parent output. Solution:

✅ Parent agent must set Output Key field
✅ Child agent must set Include Contents field with same key
✅ Agents must be in Sequential flow (not Parallel)

Agent System Overview

Learn agent types and architecture

Tools Development

Create custom tools

MCP Integration

Connect external platforms

🎯 Summary

You now know how to: ✅ A. Route to sub-agents by mentioning agent names in prompts ✅ B. Call tools & MCP by listing available functions ✅ C. Query RAG Datastores by mentioning knowledge domains ✅ D. Pass data with output keys between agents ✅ E. Apply flexible patterns with conditional logic Next Steps:

Open Admin UI → Agent Management
Create a simple Orchestrator with 2 sub-agents
Test routing with different user inputs
Gradually add tools, MCP, and RAG

Remember: No code changes needed - just configure prompts and save! 🚀

Agent Orchestration Overview Squad Collaboration

​Prompt Engineering for Agent Orchestration

​🎯 Overview

​🧠 Core Concepts

​How Junis Routing Works

​Where to Configure

​🔀 A. Sub-Agent Routing

​How It Works

​✅ Working Example (From Production)

​📋 Pattern Breakdown

​🛠️ How to Set Up

​Step 1: Create Sub-Agents

​Step 2: Link to Orchestrator

​Step 3: Configure Orchestrator Prompt

​✨ Advanced Routing Patterns

​Multi-Language Routing

​Priority-Based Routing

​🛠️ B. Tool Calls & MCP Integration

​How It Works

​✅ Working Example (Tool Call)

​📋 Pattern Breakdown

​🔧 How to Configure Tools

​Step 1: Register Tool in Admin UI

​Step 2: Add to Agent Config

​Step 3: Update Instruction Template

​✨ MCP Integration Example

​🔧 How to Configure MCP

​Step 1: Enable MCP Platform

​Step 2: Link MCP to Agent

​Step 3: Update Instruction Template

​📚 C. RAG Datastore Queries

​How It Works

​✅ Working Example

​📋 Pattern Breakdown

​🗂️ How to Set Up RAG

​Step 1: Create DataStore

​Step 2: Add RAG Tool to Agent

​Step 3: Update Instruction Template

​⚠️ Important Notes

​🔄 D. Output Format & Key Passing

​How It Works

​✅ Working Example (Parent Agent)

​✅ Working Example (Child Agent)

​📋 Pattern Breakdown

​🔧 How to Configure

​Step 1: Parent Agent Setup

​Step 2: Child Agent Setup

​Step 3: Link Agents

​✨ Advanced: Schema Validation

​🎭 E. Flexible Application Patterns

​Conditional Logic

​Multi-Step Workflows

​Dynamic Tool Selection

​Error Handling

​✅ Best Practices

​1. Be Explicit

​2. Provide Examples

​3. Use Clear Formatting

​4. Test Incrementally

​5. Monitor & Iterate

​🐛 Troubleshooting

​Agent Not Routing

​Tool Not Called

​MCP Call Fails

​RAG Returns No Results

​Output Key Not Passed

​📖 Related Documentation

Agent System Overview

Tools Development

MCP Integration

​🎯 Summary

Prompt Engineering for Agent Orchestration

🎯 Overview

🧠 Core Concepts

How Junis Routing Works

Where to Configure

🔀 A. Sub-Agent Routing

How It Works

✅ Working Example (From Production)

📋 Pattern Breakdown

🛠️ How to Set Up

Step 1: Create Sub-Agents

Step 2: Link to Orchestrator

Step 3: Configure Orchestrator Prompt

✨ Advanced Routing Patterns

Multi-Language Routing

Priority-Based Routing

🛠️ B. Tool Calls & MCP Integration

How It Works

✅ Working Example (Tool Call)

📋 Pattern Breakdown

🔧 How to Configure Tools

Step 1: Register Tool in Admin UI

Step 2: Add to Agent Config

Step 3: Update Instruction Template

✨ MCP Integration Example

🔧 How to Configure MCP

Step 1: Enable MCP Platform

Step 2: Link MCP to Agent

Step 3: Update Instruction Template

📚 C. RAG Datastore Queries

How It Works

✅ Working Example

📋 Pattern Breakdown

🗂️ How to Set Up RAG

Step 1: Create DataStore

Step 2: Add RAG Tool to Agent

Step 3: Update Instruction Template

⚠️ Important Notes

🔄 D. Output Format & Key Passing

How It Works

✅ Working Example (Parent Agent)

✅ Working Example (Child Agent)

📋 Pattern Breakdown

🔧 How to Configure

Step 1: Parent Agent Setup

Step 2: Child Agent Setup

Step 3: Link Agents

✨ Advanced: Schema Validation

🎭 E. Flexible Application Patterns

Conditional Logic

Multi-Step Workflows

Dynamic Tool Selection

Error Handling

✅ Best Practices

1. Be Explicit

2. Provide Examples

3. Use Clear Formatting

4. Test Incrementally

5. Monitor & Iterate

🐛 Troubleshooting

Agent Not Routing

Tool Not Called

MCP Call Fails

RAG Returns No Results

Output Key Not Passed

📖 Related Documentation

🎯 Summary