View source for AgentFlow Launcher Guide: Orchestrating Multi-Agent Workflows

{{menuAgentFlow|Orchestration}}
<pre style="background-color:#eee; color:magenta;>
# Launcher Guide: Orchestrating Multi-Agent Workflows

This guide covers how to use the **AgentFlow Launcher** to orchestrate complex multi-agent workflows, from simple agent creation to sophisticated inter-loquation pipelines.

## Launcher Components

```
Launcher (REST API + CLI)
├── Agent Pool         (registry of active agents)
├── Routing Engine     (message path selection)
├── State Manager      (lifecycle tracking)
├── Vector Broker      (semantic alignment via CLIO)
└── Timeline Sync      (causality recording)
```

---

## 1. Agent Instantiation via CLI

### **Create a New Agent**

```bash
llm-launcher.exe new "CodeAssistant" \
  --model "codellama:70b" \
  --endpoint "http://localhost:11434" \
  --persona "Senior Go Engineer" \
  --role "Writes production-safe Go code"
```

### **Create Multiple Agents**

```bash
# Create a narrative agent
llm-launcher.exe new "Narrator" \
  --model "llama3:8b" \
  --persona "Creative Storyteller" \
  --role "Generates engaging narratives"

# Create a documentation agent
llm-launcher.exe new "DocumentationAgent" \
  --model "mistral:7b" \
  --persona "Technical Writer" \
  --role "Produces clear API documentation"

# Create a test engineer
llm-launcher.exe new "TestAgent" \
  --model "codellama:70b" \
  --persona "QA Specialist" \
  --role "Writes comprehensive test suites"
```

### **List Active Agents**

```bash
llm-launcher.exe list

Output:
───────────────────────────────────────────
ID         Model           Status   Persona
───────────────────────────────────────────
code-asst  codellama:70b   active   Senior Go Engineer
narrator   llama3:8b       active   Creative Storyteller
docs-agent mistral:7b      active   Technical Writer
test-agent codellama:70b   active   QA Specialist
───────────────────────────────────────────
```

---

## 2. Single-Agent Queries

### **Direct Chat with Agent**

```bash
llm-launcher.exe chat CodeAssistant "Write a Go function to validate email addresses"

Output:
🤖 CodeAssistant:
func ValidateEmail(email string) (bool, error) {
    addr, err := mail.ParseAddress(email)
    if err != nil {
        return false, fmt.Errorf("invalid email: %w", err)
    }
    return addr != nil, nil
}
```

### **Get Agent Capabilities**

```bash
llm-launcher.exe info CodeAssistant

Output:
Name:      CodeAssistant
Model:     codellama:70b
Persona:   Senior Go Engineer
Role:      Writes production-safe Go code
Endpoint:  http://localhost:11434
Status:    active
Uptime:    2h 15m
```

---

## 3. Inter-LLM Communication (Inter-Loquation)

### **Two-Agent Dialogue**

Send a message from one agent to another:

```bash
llm-launcher.exe chat \
  --from CodeAssistant \
  --to DocumentationAgent \
  --message "Explain this Go function: func Add(a, b int) int"

Output:
🤖 CodeAssistant sends to 📖 DocumentationAgent:

> The Add function takes two integers and returns their sum.
> 
> Parameters:
>   - a (int): First operand
>   - b (int): Second operand
> 
> Returns:
>   - (int): The sum of a and b
> 
> Example: Add(3, 5) returns 8
```

### **Enable Vector Diffusion (Semantic Alignment)**

```bash
llm-launcher.exe chat \
  --from CodeAssistant \
  --to DocumentationAgent \
  --message "Explain this complex algorithm" \
  --vector-diffusion true

# This ensures that the semantic meaning is preserved across models
# CLIO embeddings align CodeAssistant's code semantics with DocumentationAgent's narrative
```

---

## 4. Orchestrated Pipelines

### **Sequential Pipeline**

Execute agents in sequence:

```bash
llm-launcher.exe launch-pipeline \
  --agents CodeAssistant,DocumentationAgent,Narrator,TestAgent \
  --input "Design and implement a distributed cache with strong consistency" \
  --output-file "outputs/20260513-cache-design.md"
```

**Execution Flow**:
1. CodeAssistant → Designs the cache system (code + architecture)
2. DocumentationAgent → Creates API documentation
3. Narrator → Writes a technical summary/narrative
4. TestAgent → Generates test suite

**Output**:
```
outputs/20260513-cache-design.md

# Distributed Cache Design
## Architecture (by CodeAssistant)
... [code and design details]

## API Reference (by DocumentationAgent)
... [documentation]

## Executive Summary (by Narrator)
... [narrative overview]

## Test Suite (by TestAgent)
... [test cases]
```

### **Parallel Consensus Pipeline**

Get multiple agents' perspectives on the same input:

```bash
llm-launcher.exe launch-consensus \
  --agents CodeAssistant,TestAgent,SecurityAgent \
  --query "Review this authentication implementation" \
  --merge-strategy highest_confidence

Output:
📊 Consensus Results:

✅ CodeAssistant (confidence: 0.96)
   "Implementation is clean and follows Go conventions. Good error handling."

✅ TestAgent (confidence: 0.94)
   "Adequate test coverage. Edge cases covered: invalid tokens, expiration."

✅ SecurityAgent (confidence: 0.98)
   "Uses bcrypt correctly. Recommend rate limiting on login endpoint."

📌 Merged Recommendation:
   "Production-ready with rate limiting addition."
```

---

## 5. REST API Endpoints

### **List Agents**

```bash
curl http://localhost:8080/api/v1/agents

Response:
{
  "agents": [
    {
      "id": "code-asst",
      "name": "CodeAssistant",
      "model": "codellama:70b",
      "status": "active",
      "personification": {
        "persona": "Senior Go Engineer",
        "role": "Writes production-safe Go code"
      }
    },
    ...
  ]
}
```

### **Create Agent (REST)**

```bash
curl -X POST http://localhost:8080/api/v1/agents \
  -H "Authorization: Bearer [PAT_FROM_WIN11]" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Narrator",
    "personification": {
      "persona": "Creative Storyteller",
      "prompt": "Generate engaging narratives about technology"
    },
    "llm_config": {
      "model": "llama3:8b",
      "endpoint": "http://localhost:11434"
    }
  }'

Response:
{
  "id": "narrator",
  "name": "Narrator",
  "status": "active",
  "created_at": "2026-05-13T02:19:46Z"
}
```

### **Inter-Agent Chat (REST)**

```bash
curl -X POST http://localhost:8080/api/v1/agents/{from}/chat \
  -H "Content-Type: application/json" \
  -d '{
    "to": "DocumentationAgent",
    "message": "Explain this code in plain English",
    "vector_diffusion": true
  }'

Response:
{
  "id": "msg-20260513-001",
  "from": "CodeAssistant",
  "to": "DocumentationAgent",
  "response": "This code validates email addresses...",
  "vector": [0.23, -0.67, 0.89, ...],
  "confidence": 0.98,
  "latency_ms": 234
}
```

### **Launch Pipeline (REST)**

```bash
curl -X POST http://localhost:8080/api/v1/pipelines \
  -H "Content-Type: application/json" \
  -d '{
    "agents": ["CodeAssistant", "DocumentationAgent", "Narrator"],
    "input": "Write a story about a spaceship",
    "output_format": "markdown"
  }'

Response:
{
  "pipeline_id": "pipe-20260513-001",
  "status": "running",
  "progress": "0/3 agents completed",
  "created_at": "2026-05-13T02:19:46Z"
}

# Poll for results
curl http://localhost:8080/api/v1/pipelines/pipe-20260513-001

Response:
{
  "pipeline_id": "pipe-20260513-001",
  "status": "completed",
  "progress": "3/3 agents completed",
  "output": "# Story: The Spaceship\n... [full output]"
}
```

---

## 6. Agent Dialogue Contracts

Define reusable agent communication patterns in YAML:

```yaml
# dialogues/code-to-docs.yaml
pipeline:
  source: CodeAssistant
  destination: DocumentationAgent
  trigger: "Function implemented"
  payload:
    vector_diffusion: true
    payload_type: "code"
    
  # Transformation rules
  transform:
    - extract_docstring: true
    - preserve_semantics: true
    - add_examples: true
```

Use the contract:

```bash
llm-launcher.exe dialogue \
  --contract "code-to-docs" \
  --input "func Download(url string) ([]byte, error)"
```

---

## 7. Emergence Detection & Zenkin Moments

Monitor multi-agent interactions for sudden insights:

```bash
llm-launcher.exe monitor-emergence \
  --agents CodeAssistant,SecurityAgent,ArchitectAgent \
  --query "Design a zero-trust authentication system"

Output:
⚡ Zenkin Flash Detected (entropy: 0.02, confidence: 0.98)
   Time: 2026-05-13 02:22:15
   Pattern: CodeAssistant → SecurityAgent → ArchitectAgent (3 exchanges)
   
📌 Emergence: Hybrid approach combining OAuth 2.0 + hardware keys
   First suggested by: SecurityAgent
   Refined by: CodeAssistant
   Architecture validated by: ArchitectAgent
```

---

## 8. Timeline & Causality

Track all agent interactions:

```bash
llm-launcher.exe timeline \
  --since "2026-05-13 02:00:00" \
  --format json

Output:
{
  "events": [
    {
      "timestamp": "2026-05-13T02:19:46Z",
      "agent_from": "CodeAssistant",
      "agent_to": "DocumentationAgent",
      "vector_alignment": 0.98,
      "latency_ms": 234
    },
    {
      "timestamp": "2026-05-13T02:22:15Z",
      "type": "emergence_detected",
      "agents": ["CodeAssistant", "SecurityAgent", "ArchitectAgent"],
      "pattern": "hybrid_auth_system"
    }
  ]
}
```

---

## 9. Configuration & Customization

### **Agent Personification Override**

```yaml
# agents/custom-code-assistant.yaml
name: CustomCodeAssistant
personification:
  persona: "Expert Rust Developer"
  role: "Writes memory-safe systems code"
  system_prompt: |
    You are an expert in Rust systems programming...
    Always prioritize memory safety and zero-cost abstractions.
llm_config:
  model: codellama:70b
  endpoint: "http://localhost:11434"
  temperature: 0.3  # More deterministic
  max_tokens: 2048
```

### **Pipeline Configuration**

```yaml
# pipelines/full-development-workflow.yaml
agents:
  - CodeAssistant:
      input: "Build a TODO app in Go"
      timeout: 60000
      
  - TestAgent:
      trigger_on: "CodeAssistant.done"
      input: "Write tests for:"
      context_from: CodeAssistant
      
  - DocumentationAgent:
      trigger_on: "TestAgent.done"
      input: "Generate API docs for:"
      context_from: [CodeAssistant, TestAgent]
      
output:
  format: markdown
  save_to: "outputs/app-build-{timestamp}.md"
  include_vectors: true
```

---

## 10. Troubleshooting

### **Agent Not Responding**

```bash
llm-launcher.exe debug CodeAssistant

Output:
Agent: CodeAssistant
Status: offline
Last Heartbeat: 5m 30s ago
Endpoint: http://localhost:11434
Error: Connection refused

Fix: Ensure Ollama is running on port 11434
  ollama serve
```

### **Semantic Misalignment**

```bash
llm-launcher.exe diagnose-vector \
  --from CodeAssistant \
  --to DocumentationAgent \
  --message "Example code"

Output:
Vector Alignment: 0.73 (below threshold 0.85)
Recommendation: Increase context window or use more specific prompts
```

---

## Next Steps

1. **Create Your First Agent** – Use the CLI to create an agent
2. **Run a Simple Pipeline** – Chain two agents together
3. **Monitor Emergence** – Watch for Zenkin moments in multi-agent dialogue
4. **Build Custom Contracts** – Define reusable agent patterns

See [ARCHITECTURE.md](ARCHITECTURE.md) for design details and [API_SPEC.md](API_SPEC.md) for full API reference.
</pre>