Autonomous Agent Orchestration (Nova)
CrewAI agency with CEO, 5 divisions and 29 specialists running real work 24/7 — FastAPI, 3D graph, SSE traces and local inference. This isn't a chatbot with a nice name: it's a business operated by AI.
Nova MWS is a real orchestration product, not a chatbot with a nice name. It coordinates a network of 29 specialist agents (research, content, ops) with per-agent config, live traceability and a direct bridge into MWS business data.
Real screenshots of the system in operation.
Problem
- 1Cut manual ops without losing control.
- 2Keep privacy with local inference when needed.
- 3Make a 30-role network observable.
- 4Ship reproducibly (Docker / VPS).
The approach
Multi-agent design
CEO + divisions + 29 specialists with explicit routing and runtime overrides.
API + ops panels
FastAPI with /visual, /vivo (SSE), /configuracion and health.
Local inference & tools
Ollama plus MCP/Composio/MWS tools for real work.
Observability
Run traces, optional Langfuse, module status in config.
Partial control-plane map: operator and live panel → orchestration API → director hub / divisions / specialists → local inference, bounded tools and state. No private configs or full agent inventory.
System layers
- 01Ops UI: /visual, /vivo, /configuracion
- 02FastAPI: run, SSE stream, graph, health
- 03CrewAI orchestration: CEO → hubs → specialists
- 04Inference: Ollama + per-agent overrides
- 05Tools: MCP, Composio, MWS pack, n8n
- 06Data: SQLite/PostgreSQL + runtime overrides
Applied stack
- Python
- CrewAI
- FastAPI
- Ollama
- RAG
- MCP
- Docker
- SSE
Design decisions
Local-first vs cloud-only LLMs
Privacy, predictable cost, customer perimeter control.
Specialist network vs one assistant
Domain routing and clear capability ownership.
Per-agent config without redeploy
Operate prompts/models/tools from a panel, not git every time.
Impact
- 29+1
- Specialists + CEO orchestrated
- 12.2M
- Tokens and 296 crew calls for ~USD 29.90 (14-provider cascade)
- 2,749
- RAG chunks indexed over 159 first-party documents
- 11
- Subsystems monitored in the health HUD
- 24/7
- Continuous ops with traces
- Local
- Private inference via Ollama
Key points
- Ships with visual panel and live execution—not a notebook demo.
- Local-first architecture with controlled external tools.
- End-to-end ownership: agents, API, ops UI, deploy.
Fit / how to hire
Hire this if you need a Solutions Architect / AI automation lead who has already proven a 29-agent ecosystem can run in production, not just in a notebook. We start with a 15-min call; if scope calls for it, I stay on as an ongoing retainer.
Book a technical call
