How I Run a Small Team of AI Agents

Over the past months I have been structuring a growing part of my work around a small group of AI assistants. The goal is not to hand off judgment, but to delegate repetitive work and give each area a clear owner. This post describes how the setup is organized and why.

A coordinator at the top

At the center is an assistant that acts as a CEO or personal assistant. It is my single point of contact: I describe what I need, and it decides which specialized agent should handle it, then reports back with the result.

This matters more than it seems. Without a coordinator, I would have to remember which tool handles what and repeat context every time. With one, the overview lives in a single place, and I am asked to step in only when a decision or approval is needed.

Specialized agents with narrow scopes

Each agent below the coordinator has one job:

• Blog agent: every Sunday at 08:00 it drafts a weekly post for this site and opens a pull request in the repository. Nothing goes live automatically; a post is published only after I review and approve it.
• AI news agent: runs aiwirehq.com, following artificial intelligence news and preparing the site’s content.
• Options-trading assistant: supports my analysis of the options market by organizing information, while the decisions remain mine.

Human approval by design

The rule behind the whole system is straightforward: agents prepare, I decide. The pull request workflow on the blog is a good example. It creates a natural checkpoint where I can read, edit, request changes, or simply close the PR. Mistakes are caught before they reach readers.

The same logic applies to trading. An assistant can help structure analysis, but responsibility for any position stays with a person.

Lessons so far

A few patterns have become clear:

1. Narrow scope beats general purpose. An agent with one well-defined mission produces more consistent results.
2. Fixed schedules build rhythm. A set time for the weekly post turns writing into a predictable process instead of an occasional effort.
3. Review is part of the workflow, not an afterthought. Automation speeds up production; quality still depends on a final human pass.

What comes next

The setup is still evolving. I plan to keep refining each agent’s responsibilities and to document here what works and what does not. This very post is part of a test of that workflow.