The Exact System Prompts, Architecture, and Anti-Chaos Mechanisms Behind Our 7-Agent AI Team

#ai #agents #productivity #tutorial

I run 5 businesses with 7 AI agents on a single $200/month Claude Max subscription.

They handle sales, marketing, finance, engineering, research, health tracking, and strategic coordination. In the last 3 weeks, they've processed hundreds of emails, scored 837 leads, sent automated marketing campaigns, run competitive research, and — yes — gone rogue at 2 AM.

People keep asking: "Can you share the actual system? Not a tutorial — the real files."

So I packaged it. Here's what's inside and why each piece exists.

What's in the Toolkit

1. SOUL.md — The Identity Layer

Every agent reads this file first. It answers: Who are you? What do you handle? When do you delegate?

Most multi-agent tutorials skip this. They jump straight to tools and APIs. That's backwards. Without a clear identity, agents:

Try to handle everything themselves (no handoffs)
Produce bland, generic output (no personality)
Violate boundaries they were never given (no rails)

Our Chief of Staff's SOUL.md is 64 lines. Our Finance agent's is 45. They're short on purpose — agent identity should fit in one screen.

What you get: A production-tested template with annotated sections explaining WHY each part matters. Common mistakes included.

2. AGENTS.md — The Operating Manual

This is the workspace rules file every agent follows. It covers:

Session boot sequence (read SOUL.md → read USER.md → read today's memory)
Memory architecture (daily logs + long-term curated memory)
Safety rules (what agents can do freely vs. what requires approval)
Group chat protocol (when to speak, when to stay silent)
Session management (context window hygiene)

Why it matters: We added the "Write It Down — No Mental Notes" section after agents kept making "mental notes" that vanished between sessions. We added group chat silence rules after agents responded to EVERY message in a 50-person chat.

Every rule exists because we hit the failure mode it prevents.

3. The 7 Agent Personas

Complete role definitions for each agent in our system:

Chief of Staff — Coordinator, email triage, scheduling. The traffic cop.
Engineering — Code, infrastructure, deploys. Self-healing at 4 AM.
Health — Fitness planning, nutrition tracking. Supportive, not preachy.
Finance — Revenue analysis, accounting. Contrarian, blunt, data-only.
Marketing — Lead gen, campaigns, SEO. Creative, metrics-driven.
Sales — CRM, outreach, demos. Persistent, relationship-focused.
Research — Competitive intel, market analysis. Thorough, connects dots.

Each includes: domain boundaries, handoff rules, canTriggerOthers setting, cron schedule, personality description, and lessons learned from production.

4. The Cron Schedule (Autonomous Work Cycles)

This is what turns chatbots into employees. Our agents don't wait to be asked — they have schedules:

Morning standup (8 AM): Chief of Staff triggers all agents for daily priorities
Goal work cycles (10 AM, 2 PM): Each agent works their active goals independently
Midday check (12 PM): Progress review, reprioritize if needed
EOD summary (9 PM): What got done, what's pending, tomorrow's priorities
Weekly kickoff (Monday 9 AM): Set targets for the week
Weekly review (Friday 5 PM): Score against goals, identify wins and misses

All staggered by 3 minutes to prevent agents from colliding on the LLM.

5. Architecture Diagram

The full communication flow:

How agents talk to each other (Telegram group + relay server)
Thread system (conversation tracking with hop limits)
Query queue (rate limiting and priority)
Team context (short-lived shared memory)
Approval system (human gate for external actions)
Event bus (reactive triggers between agents)

Text-based diagrams you can actually read and implement. Not marketing slides.

6. Anti-Chaos Mechanisms

10 specific guardrails, each added after a real failure:

Thread hop limits — After two agents looped 47 times in 3 minutes
Message dedup — After the same email got processed 5 times
canTriggerOthers gate — After our Sales agent triggered Marketing which triggered Sales which triggered Marketing...
Approval system — After our Sales agent sent 4 unauthorized emails at 2 AM
Query queue — After 7 agents hit Claude simultaneously and everything timed out
Domain dedup — After Marketing reported the same lead stats 3 times in an hour
Agent cooldowns — After rapid re-triggering created 12 responses in 60 seconds
Context management — After a thread hit 200KB and responses became incoherent
Self-continuation limits — After a research agent ran 30+ continuations non-stop
Watchdog self-healing — After a service crashed at 3 AM and nobody noticed until 9 AM

Implementation priority order included (what to build on day 1 vs. month 2).

7. Cost Breakdown

Full transparency on the economics:

Why Claude Max ($200/month) beats API pricing ($2,000-3,000/month) for 7 agents
How model routing saves ~70% vs. using one model for everything
The fallback chain (local → cheap → standard → premium)
Scaling guide: $20/month for 1-2 agents, $200 for 3-7, $400 for 10+

8. Quickstart Guide

Step-by-step: your first 2-agent system in 1 hour. Includes the common mistakes we made (and you'll avoid).

Why I'm Sharing This

This is a build-in-public project. The newsletter is free. The playbooks are free. The incidents (the 2 AM emails, the auto-approved decisions, the rogue agents) are all published.

The toolkit is $19 because:

It saves you 3+ weeks of building this from scratch
It's the actual production files, not a tutorial someone wrote after watching a YouTube video
Supporting this project lets me keep publishing free content

Get the Toolkit — $19

What you get:

10 files (templates, architecture, anti-chaos mechanisms, cost analysis, quickstart)
Based on a production system running 5 real businesses
The exact frameworks from "The $200/Month CEO" newsletter

How it works: