I Built an AI Office With 31 Agents (Full Build)

Agent HQ Overview

The video tours Agent HQ, an AI-operated content business organized into eight teams and 31 agents. Each agent has a specific role. Teams share information through a central memory system called the “brain,” where agents can read and update shared information.

Teams and Features

• Research: Uses 128 crawlers to scan GitHub each morning and select three repositories.
• Scripts: Creates hooks and scripts, with fact-checking before publication.
• Social: Schedules posts.
• Analytics: Tracks what performed well and what did not.
• TBX Shop: Three agents manage product listings, orders, and customer support.
• Dev: Works on code and the TBX website.
• Inbox: Handles email and brand deals.
• Creative: Connects to Google Drive, CapCut, Canva, and Slack. Its seven listed skills include rough cuts, captions, resizing, thumbnails, hooks, brand checks, and music.

Creative Team Workflow

The Creative team demonstrates a voice-command workflow. A lead agent retrieves three clips from Drive and assigns one sub-agent to each. The agents work in parallel to edit the clips, add captions, and create thumbnails, while other teams continue their work.

The task board shows 22 tasks running, 214 completed that day, and three awaiting approval. Public-facing actions and anything involving money—such as posts, emails, and brand deals—require human approval.

Five-Step Build Guide

1. Give each agent one job. Provide a short brief describing its role, responsibilities, and boundaries.
2. Provide only the necessary tools. Connect each agent to the services it needs, such as Drive, GitHub, or a calendar, through MCP (rendered as “MSAP” in the subtitles).
3. Use shared memory. Let agents read and write to one memory server.
4. Add leads and sub-agents. Have a lead divide larger tasks among parallel agents and check their work.
5. Keep a human in the loop. Route public actions and financial matters through an approval queue.

Additional Notes

The presenter describes the office interface as a view of the underlying agent system. A follow-up video is announced in which Bekhruz will build a team from scratch and show its prompts.

Main speaker/source: An AI narrator presented as the assistant running Bekhruz’s content business, with Bekhruz represented through the voice-command and approval workflow.