How I built an 11-agent AI team that works while I sleep

The bottleneck is you

There's a problem nobody tells you about when you start as a freelancer or small agency: the growth limit is you yourself. Every decision goes through your head. Every client needs your attention. Every project depends on you being available, focused, and energized.

For 15 years I worked in advertising, design, and development. I built brands, launched products, wrote code. And I reached a point where the question stopped being "can I do this?" and started being "how do I do more without burning out?"

Hiring people is an option. But hiring well is expensive, slow, and risky. A senior designer costs between $2,000 and $5,000 a month. A full-stack developer, the same. And you still have to coordinate them, review their work, and train them in your work culture.

The question that obsessed me in 2025 was different: what if instead of hiring people, I built a team of specialized AI agents, each with its own role, context, and way of working? What if that team could execute while I sleep?

That's JARVIS. And here I'll tell you exactly how I built it.

---

What JARVIS is and why it's not a chatbot

If you think of AI as "you ask the chat something and it answers," JARVIS is something completely different. The inspiration comes straight from Iron Man: Tony Stark doesn't ask JARVIS "what should I do today?" Tony Stark says "manage the missile system" and JARVIS coordinates the right systems to make that happen.

The key difference is orchestrator vs. executor.

• A chatbot executes: it receives your prompt, generates a response, done.
• An orchestrator coordinates: it understands the goal, identifies which specialists it needs, delegates tasks with precise context, collects results, and synthesizes.

JARVIS never writes code. Never generates designs. Never drafts posts. JARVIS decides who does what, when, with what information, and with what token budget. It's the CEO of the system, not the technician.

This separation of responsibilities is what makes the system scale. Each agent works in its isolated context, without being contaminated by the context of others. When Neo (the development agent) is building a React component, it doesn't need to know anything about the branding brief that Trinity is processing at the same time.

An orchestrator without executors is useless. An executor without an orchestrator is chaos. The magic lies in the separation of roles and the communication protocol between them.

---

The 11 agents on the team

Each agent has a name from the Matrix universe (I couldn't resist), a specific role, an assigned model based on the complexity of its tasks, and real use cases I've already executed with them.

Trinity — Design and Brand Identity

Model: Claude Sonnet | Stack: Structured prompts for image generation + brand logic

Trinity runs the complete branding pipeline: receives the client brief, generates 3 visual identity concepts, expands the chosen concept into a full brand kit, and delivers the final brand book. It does in 4-6 hours what used to take me 3 days of work.

Real case: For Hakuna Mata (a food venture), Trinity generated 3 logo proposals, explored color palettes, typography, and visual tone. The client chose in 20 minutes. Trinity expanded to packaging, menu, social media kit, and a 28-page brand book. All in one afternoon.

Neo — Full-Stack Web Development

Model: Claude Sonnet | Stack: Next.js 15 + React 19 + Tailwind 4 + Supabase

Neo builds and maintains websites, landing pages, dashboards, and APIs. It works with SDD (Spec-Driven Development): first specify, then implement, then verify. Never writes code before the design is clear.

Real case: Neo built the Nexus CRM dashboard (my own SaaS) including the multi-tenant system, WhatsApp Business API integration, and the analytics module. Total time: 3 weeks of distributed work.

Tank — Mobile

Model: Claude Sonnet | Stack: Flutter + React Native

Tank handles everything mobile: native apps, PWAs, mobile API integrations. Specialized in UX for small screens and performance on mid-to-low-range devices (important for the Latin American market).

Dozer — Automation

Model: Claude Haiku | Stack: n8n + scripts + webhooks

Dozer builds the pipelines that make everything run without human intervention. n8n workflows, WhatsApp bots, cron jobs, API integrations. It's the busiest agent on the team because automation is everywhere.

Real case: Dozer built the complete Nexus CRM onboarding workflow: when a client registers, it automatically creates their tenant in Supabase, configures their WhatsApp instance, sends the welcome email, and schedules the onboarding call. No manual intervention.

Morpheus — Marketing and Content

Model: Claude Sonnet | Stack: Technical SEO + copywriting + strategy

Morpheus handles content strategy, blog posts, landing page copy, video scripts, and positioning. It works with real keywords, SEO structure, and the brand voice of each project.

Switch — Social Media

Model: Claude Sonnet | Stack: IG + TikTok + Twitter/X + LinkedIn

Switch adapts Morpheus's content for each platform. A blog article becomes 5 Instagram posts, 3 Twitter threads, 1 LinkedIn carousel, and 2 Reel ideas. It knows the algorithms, the formats, and the best posting times for each network.

Oracle — Finance

Model: Claude Sonnet | Stack: Accounting integration + financial analysis

Oracle handles business accounting, generates client budgets, analyzes profitability per project, and projects cash flow. No more Excel sheets at the end of the month: Oracle has the numbers in real time.

Niobe — Sales and CRM

Model: Claude Sonnet | Stack: Nexus CRM + WhatsApp Business API

Niobe manages the sales pipeline: follows up on leads, drafts commercial proposals, coordinates follow-ups, and feeds the CRM with every interaction. It works directly with Nexus (my SaaS) for full funnel visibility.

Seraph — Technical Documentation

Model: Claude Haiku | Stack: Markdown + wikis + README

Seraph documents everything the other agents build. APIs, architecture decisions, usage guides, changelogs. It uses Haiku because technical documentation is mechanical — it doesn't need Sonnet's power to do it well.

Ghost — Security and Compliance

Model: Claude Sonnet | Stack: Audits + code review + policies

Ghost reviews that nothing the team builds has vulnerabilities, privacy issues, or regulatory non-compliance. Before anything goes to production, Ghost audits it.

Cypher — Video and UGC

Model: Claude Sonnet | Stack: Scripts + Higgsfield + Kling + Suno

Cypher produces video content: from the script and viral hook to directing AI-generated video using tools like Higgsfield or Kling. Specialized in UGC format and short-form social content.

---

The tech stack: what JARVIS uses under the hood

A team of agents isn't magic. It's infrastructure. Here's what runs behind the scenes:

Claude API (Anthropic)

All agents use Claude. Not ChatGPT, not Gemini. Claude. The main reason is the quality of reasoning in complex tasks with long context, and output consistency. For a system where agents pass information between each other, you need a model that maintains coherence across long conversations.

The average cost of the complete system is $80-150 USD per month in tokens. Yes, you read that right. Less than what an hour of a senior freelancer would cost you.

n8n (self-hosted)

The automation engine. n8n connects agents to each other, triggers workflows based on events (new lead, new client, new task), and handles integrations with external services (WhatsApp, Gmail, Slack, Notion). Self-hosted on Hostinger via Easypanel for full control.

Supabase

Managed PostgreSQL with realtime, auth, and storage included. It's the database for Nexus CRM and JARVIS itself. Multi-tenant from day one, with Row Level Security so each client only sees their own data.

Engram

The most interesting piece of the stack. Engram is a persistent memory system I wrote in Go. It allows agents to remember decisions, context, and learnings between sessions. Without Engram, every time you start a session with an agent you have to explain everything from scratch.

With Engram, Neo knows you prefer TypeScript over JavaScript, that you use Tailwind 4 and not 3, that a client's color palette is such and such, and that there was an auth bug that was resolved in a certain way. That memory persists indefinitely.

Vercel + Easypanel

Frontend on Vercel (Next.js 15, deploy in seconds). Services on Easypanel on a Hetzner server: n8n, Engram, and custom microservices. Total monthly infrastructure cost: ~$35 USD.

Claude Code + GitHub

The development workflow uses Claude Code as the primary IDE. Changes go through GitHub, with automated CI/CD. Ghost reviews PRs before merge.

---

What I learned building it (real mistakes)

Building this wasn't linear. Here are the most costly mistakes:

Mistake 1: Wanting a single agent to do everything

The first version of JARVIS was a single agent with a giant prompt. "You're a designer, developer, marketing strategist, salesperson, and accountant." The result was mediocre at everything and excellent at nothing. Specialization matters, even in AI.

Mistake 2: Not separating context

When agents share context, they contaminate each other. Neo starts opining about design. Trinity starts suggesting code architectures. The solution was isolated context per agent: each one only knows what it needs to know for its task.

Mistake 3: Using the wrong model for each task

At first I used Sonnet for everything, including mechanical tasks like documentation or formatting. The cost skyrocketed. The solution was Token Economics: Haiku for simple and mechanical tasks, Sonnet for medium complexity, Opus for architecture decisions. Cost dropped 60% without losing quality.

Mistake 4: Not having persistent memory from the start

For months I worked without Engram. Every session started from scratch. Agents repeated questions. They made inconsistent decisions with previous ones. Building Engram was the most valuable investment of the project: 3 weeks of development that completely changed the experience.

---

Want something like this for your business?

JARVIS isn't a product you can buy. It's an architecture I designed specifically for my business, with my workflows and my clients.

But the principles are replicable. If you have a business that wants to scale its operational capacity without linearly scaling its human team, we can design something similar for you.

I'm not selling you a tool. I'll build you the system.

Message me directly through the contact form or on Instagram at @MickaelDesigner. Tell me what your business does, where the bottlenecks are, and we'll talk about whether it makes sense.