https://reddit.com/link/1rh95cl/video/0b8dqf83x9mg1/player

The setup that shouldn't work but does

I have 13 AI agents that work on marketing for my product. They run every 15 minutes, review each other's work, and track everything in a database.

When one drafts content, others critique it before I see it. When someone gets stuck, they ping the boss agent. When something's ready or stuck, it shows up in my Telegram.

It's handling all marketing for Fruityo (my AI video generation platform).

Here's the architecture and how you could build something similar.

The problem

Most AI workflows are single-shot: ask ChatGPT → get answer → copy-paste → lose context → repeat tomorrow.

That works for quick questions. It breaks down for complex work that needs:

• Multiple steps across days
• Research that builds on previous findings
• Different specialized perspectives (writing vs strategy vs critique)
• Quality review before anything ships
• Tracking what's done, what's blocked, what's next

I needed AI that works like a team, not a chatbot, and I saw some guys on Twitter building UI's for OpenClaw agents...

The architecture

Infrastructure:

• OpenClaw - gives agents the ability to browse the web, execute commands, manage files, and interact with APIs
  • Cron - schedules agent heartbeats
  • Telegram - notification layer (agents ping me when something needs attention)
• PocketBase - database storing tasks, comments, documents, activity logs, goals
• Claude Max

Workflow: Tasks move through states: backlog → todo → in_progress → peer_review → review → approved → done

Each state has gates. Agents can't skip peer review. Boss can't approve without all reviewers signing off. I'm the only one who moves tasks to done.

The team (from Westeros)

Each agent has a role, specialty, and personality defined in their SOUL.md file:

Agent	Role	What they do
🐺 Jon Snow	Boss	Creates tasks, coordinates workflow, and promotes peer-reviewed work to final review
🍷 Tyrion	Content Writer	Writes tweets, threads, blog posts, landing pages in my tone.
🕷️ Varys	Researcher	Web research, competitor analysis, data mining
🐉 Daenerys	Strategist	Campaign planning, positioning, and goal setting
⚔️ Arya	Executor	Publishes content, runs automation, ships work
🦅 Sansa	Designer	Creates design briefs, visual concepts
🗡️ Sandor	Devil's Advocate	Gives brutal, honest feedback, catches BS
...	...	...

Why Game of Thrones names? Why not, I love GOT :) ...and personality matters. Sandor reviews content like a skeptic. Tyrion writes with wit. Varys digs for hidden data. Their SOULs define behavior - Sandor will roast bad writing, Daenerys will flag strategic misalignment.

Better to have multiple specialists with distinct viewpoints than one mediocre generalist.

How it actually works: The heartbeat protocol

Each agent has its own OpenClaw workspace. Every agent runs a scheduled heartbeat every 10 minutes (scattered by 1 minute each to avoid hitting the DB simultaneously).

What happens in a heartbeat:

1. Agent authenticates, sets status to "working"

Connects to PocketBase, updates the status field so others know it's active.

2. Reviews others FIRST (highest priority)

• Fetches tasks where other agents need my review
• Reads task description, existing comments, documents they created
• Posts substantive feedback (what's good, what needs fixing)
• If work is solid → leaves approval comment
• If needs changes → explains exactly what's wrong

This is the peer review gate. If I'm assigned to the same goal as you, I MUST review your work before it moves forward.

3. Works on own tasks

• Fetches my assigned tasks from DB
• Picks up anything in todo → moves to in_progress
• Does the actual work (research, write, analyze, etc.)
• Saves output to PocketBase documents table
• Posts comment explaining approach
• Moves task to peer_review (triggers all teammates on that goal to review)
• Logs activity to activity table

4. Updates working status, sets to "idle"

Agent writes progress to PROGRESS.md (local state tracking), sets PocketBase status to "idle", waits for next heartbeat.

Task Flow Example

Goal: Grow Fruityo on socials

Jon creates the task to create a post about current UGC video trends and assigns it to Varys (researcher). I approve it by moving from backlog to todo.

Varys picks it up, moves to in-progress, researches, saves findings to the database, and moves to peer review. Daenerys and Tyrion review his work, suggest improvements. Varys creates new version based on feedback. Once both approve, Jon (boss) promotes the task to the review stage.

I get a Telegram notification, review the research document, and approve. Task moves to done.

All communication happens via comments on the task. All work is stored in the database. Context persists.

The boss role: Why Jon is special

Jon isn't just another agent. He has special authority:

Only Jon can:

• Create new tasks (via scheduled cron, analyzing goals)
• Promote tasks from peer_review → review (after all peers approve)
• Reassign tasks when someone's blocked
• Change task priorities

Jon's heartbeat is different:

• Checks if peer_review tasks have all approvals → promotes to review
• Identifies blocked tasks (stuck over 24 hours) → investigates why → escalates to me
• Coordinates handoffs between agents

Think of it like: agents are the team, Jon is the team lead, and I am the executive.

Without a coordinator, you'd have chaos - 7 agents all trying to assign work to each other with no one having the final word.

Goals: How work gets organized

Here's where it gets interesting. Instead of creating tasks manually every day, I define long-term goals and let Jon generate tasks automatically.

A goal defines:

• What we're trying to achieve
• Which agents are assigned to it
• How many tasks should Jon create per day/week

Example: I created a goal "Grow Fruityo twitter presence." Assigned agents: Varys (research), Tyrion (writing), Arya (publishing), Sandor (review). Told Jon to create 3 tasks per day related to this goal.

Every day, Jon analyzes the goal, 15-day tasks history, creates 3 relevant tasks in the backlog ("Research trending AI video topics," "Draft thread on B-roll generation," etc.), and assigns them to the right agents. And I edit and/or just move good ones to todo.

Why this matters:

1. Selective peer review - Only agents assigned to that goal review each other's work. I can have 20+ agents in the system, but only the 4 assigned to "Twitter content" review those tasks. Saves tokens, keeps review relevant.
2. Automatic task generation - I set a goal once, Jon creates tasks daily/weekly. No manual planning every morning.
3. Scope control - Different goals can have different agent teams. Marketing goals get Tyrion/Varys/Arya. Product goals get different specialists.

You could run multiple goals simultaneously - each with its own team, its own task cadence, its own review process.

Communication Layer

All agent communication happens through PocketBase comments on tasks.

To reach another agent → mention their name in a comment
To reach me → mention my name in a comment (notification daemon forwards to Telegram)
To reach Jon specifically → dedicated Telegram topic (thread) bound to Jon's OpenClaw topic

No DMs, no scattered Slack threads. Everything on the task, in context, persistent.

What I use it for

HQ runs almost all marketing for Fruityo:

- Competitor research
- Reddit research
- Twitter threads
- Blog posts
- Landing page copy
- Campaign planning
- Design briefs
- Content publishing (soon)
- ...Whatever agents have skills for

Before: I'd spend 1 day per blog post (research, draft, edit, publish)
With HQ: ~30 minutes of my time to review and approve. Agents handle research, drafting, peer review.

The quality is better because of peer review. Varys catches bad data. Daenerys catches strategic drift. Sandor catches AI clichés and marketing BS.
> YES, this could burn through tokens quite quickly (safu on Claude Max sub), but it seems, that I found the right combination of setup and context optimisations.

If you want something similar

This is my custom setup, built for my specific needs. But the pattern is generalizable - you could use it for content creation, product development, research projects, or any work that needs multiple specialized perspectives with quality gates.

• All of this is built on OpenClaw (open source AI agent framework)
• PocketBase is free and self-hostable
• FULL GUIDE above is free. Just prompt your little lobster the right way :)

If you build something like this, I'd love to hear about it. Reply with what you'd use it for or what you'd do differently.

If you'd like to see this packaged as a ready-to-use product or like to know even more details, let me know here.