Agent Swarm Development System

Sistema de desarrollo para crear infraestructura de agentes usando:

• Anthropic Swarm: Patrón de orquestación de agentes
• Synapse: Memoria neuro-simbólica (pmaojo/synapse-engine)
• MCP: Model Context Protocol para herramientas
• Vercel: Despliegue

Estructura

agent-swarm-dev/
├── SKILL.md              # Esta skill
├── SPEC.md               # Especificación del sistema
├── agents/               # Agentes especializados
│   ├── orchestrator.md   # Coordina flujo (Anthropic Swarm)
│   ├── coder.md         # Genera código
│   ├── memory.md         # Memoria Synapse
│   └── reviewer.md       # Revisa calidad
├── scripts/
│   ├── init_swarm.sh     # Iniciar proyecto
│   ├── run_agent.sh      # Ejecutar agente
│   ├── deploy.sh         # Desplegar a Vercel
│   ├── synapse_agent.py  # Tool Python SDK
│   └── synapse_mcp.py    # Servidor MCP
├── .mcp/
│   └── config.json       # Configuración MCP
└── deploy/
    └── vercel.json       # Config Vercel

Uso

1. Iniciar Proyecto

./scripts/init_swarm.sh mi-proyecto

2. Ejecutar Agente

./scripts/run_agent.sh orchestrator "Crear una API REST"

3. Desplegar

./scripts/deploy.sh

MCP Tools (Synapse)

Tool	Descripción
query_graph	Consulta todos los triples
ingest_triple	Añade un triple
query_sparql	Consulta SPARQL
add_observation	Añade observación
ingest_memory	Ingesta múltiples triples

Flujo Swarm

User → orchestrator → (handoff) → coder → (handoff) → reviewer → (handoff) → deployer → URL
         ↑                                      ↓
         └────────── Memory (Synapse) ←────────┘

Spec-Driven Development (GSD + OpenSpec)

Two frameworks for spec-driven development are integrated:

GSD (Get Shit Done)

npx get-shit-done-cc@latest --claude --global

• /gsd:spec - Generate SPEC.md
• /gsd:build - Execute build
• /gsd:test - Run tests
• /gsd:verify - Verify implementation

OpenSpec (@fission-ai/openspec)

npm install -g @fission-ai/openspec@latest

• /opsx:new <feature> - Create new feature
• /opsx:ff - Generate full planning docs
• /opsx:apply - Implement tasks
• /opsx:archive - Archive completed

Both work with Claude Code, OpenCode, and OpenClaw via MCP.

Synapse Integration

Python SDK

from synapse import get_client
client = get_client()
client.ingest_triples([{"subject": "agent_1", "predicate": "completed", "object": "task_123"}])

MCP

El servidor MCP (synapse_mcp.py) expone las tools via JSON-RPC stdio.

GSD Integration (Get Shit Done)

Sistema de context engineering y spec-driven development. Útil para estructurar tareas complejas.

Comandos GSD

• /gsd:spec - Generar SPEC.md desde description
• /gsd:build - Ejecutar build completo
• /gsd:test - Ejecutar tests
• /gsd:verify - Verificar implementación vs spec

Flujo GSD

1. Describe lo que quieres construir
2. GSD extrae contexto y genera spec
3. Claude Code ejecuta y verifica
4. Iterar hasta correcto

Install

npx get-shit-done-cc@latest --claude --global

namespaces

• swarm: Memoria del swarm
• agents: Estado de agentes
• tasks: Tareas y resultados

🧠 Kilo-Style Interactive Mode

The swarm now includes an interactive command center for developers.

Usage

python3 scripts/kilo_interactive.py

Commands

• /ask <query> - Chat with LLM (context-aware).
• /code <task> - Run CoderAgent (e.g., "Implement login").
• /review - Run ReviewerAgent on recent changes.
• /browser <query> - Search documentation using BrowserTool.
• /harvest <path> - Scan codebase for knowledge tags.
• /scenario <name> - Load a domain-specific ontology scenario.

🔗 Smart Context & Knowledge Harvesting

The system uses advanced context parsing to reduce hallucinations and enforce consistency.

1. Smart Context (@file)

In any prompt (CLI or Agent), use @file:path/to/file to inject its content AND its associated "Golden Rules" from Synapse.
Example:

/code Refactor @file:agents/coder.py to use async/await.

2. Knowledge Tagging (@synapse)

Agents and Developers can teach the swarm by adding comments in code:

• Constraints: // @synapse:constraint Always use Pydantic v2. -> Ingested as nist:HardConstraint.
• Lessons: // @synapse:lesson Retry logic is needed for Synapse gRPC. -> Ingested as swarm:LessonLearned.

Run /harvest . or python3 agents/tools/knowledge.py to consolidate these into the Knowledge Graph.

3. Browser Tool

The CoderAgent is equipped with a headless browser (Playwright + DuckDuckGo) to:

• search_documentation(query): Find solutions online.
• read_url(url): Extract knowledge from docs.

4. Scenario Loading

Load specialized knowledge packages:

/scenario core (Loads Schema.org, PROV-O, etc.)