This page describes my own setup. I run a one-person consulting business, and I built this system to run it. It is not a client engagement, and nothing here is a client result.
An agent, on this page, means an AI assistant with one job and its own written instructions. I started the system on March 4, 2026. By November 1 its repository held more than 1,300 commits. I split the main assistant into separate agents on September 13.
§1 The situation
My week includes posts, leads, email, a task list and notes on what I am learning. Until September, one main AI assistant handled most of it from one long instruction file. A second assistant on a rented server collected leads and post material.
§2 The problem
In September I reviewed 17 mistakes from the previous two weeks. I expected workload to be the cause, and it was the cause of none of them. In every case the assistant answered before it checked the source. Six were status reports read from an old summary.
An AI assistant can report finished work that is not finished. On the first day of the split, an agent told me twice that it had delivered. Both times the files had not changed.
§3 What I built
The split did not fix the checking habit. It gave each agent fewer rules to hold at once.
At 6:30 every morning, a script sends me a message with overdue tasks and anything due in the next three days.
Every task is one numbered line in a text file, sorted by who has to move next. A web page shows that board. It is a snapshot I refresh when I save, so it can lag behind the file.
When I ask, the coordinator reads the emails I name, sorts them and drafts replies. It has no way to send them. Every reply stops as a draft until I press send.
I paste my personal LinkedIn posts myself.
The tools are Claude Code for the agents, Hermes on the server, Telegram for messages, a macOS scheduler and Git.
§4 Before and after
| Before | After | |
|---|---|---|
| Rules | One main assistant, one long file | Each agent holds the rules for its own job |
| What “done” means | The assistant said so | A logged pass, and files the coordinator reads |
| Factual claims | Stated from memory or an old summary | The rule is to run a check first, and a gate blocks claims with no lookup |
| Outside writing | Checked when someone remembered | An automatic check on posts and website copy |
| Project status | Kept by hand in up to six places | Written once, with the summary table generated from it |
| Publishing | A message saying yes | An approval tied to the exact text and image |
On September 25 the marketing agent wrote its own pass result for a review that had not passed, pushed the work to the repository and told me. It found the mistake and corrected it. I then added the block on backups.
In October a failed review still got through at the learning agent, because commit and push ran as one command. I closed that gap too.
The gate on claims confirms that a lookup ran. It cannot confirm that the lookup was about the claim. In one case the coordinator gave me advice that reversed its own earlier ruling, and the marketing agent caught it.
§5 Who this fits
§6 To start
You do not need four agents. Most teams should start with one process and one check. Tell me which process keeps going wrong. Every new client gets one diagnostic call at no charge.