All insights

Four AI agents, five working days, zero interventions

The pilot is the product. What ran here is what I now set up for others.

I wanted to know whether named AI agents with clear rules can work productively for a week without me spending more than a report read each day. So I ran my own company as the test. Every number below comes from the daily reports and the files the agents left behind.

Setup in one day

On 25 August I reviewed the security of the bot framework, set up four agent profiles (builder, content, sales and ops) and gave each one a rule file. The file says what the agent is for, which data it may read and what it must never do. I added four role playbooks, a daily routine and a first offer hypothesis. A first message between two agents went through the same day.

The rules fit on a page

  • No agent sends, posts or deletes anything.
  • Sensitive data is read only by a protected path, never by the agents themselves.
  • Every claim needs a file that proves it, which I call the truth gate.
  • The human stays the boss.

A scheduler started the round on working days. The agents ran on free-tier models in the pilot week, and heavier thinking was delegated to an existing Claude subscription when needed.

What the team delivered

  • 5 working days with output: 25, 26 and 27 August, 31 August and 1 September.
  • On all 5 days all 4 agents delivered.
  • 16 role artifacts: 5 from the builder, 5 from sales and 6 from content.
  • 5 daily reports and 6 LinkedIn drafts.
  • 0 blockers in the reports and 0 interventions from me beyond reading.

The builder documented a complete life cycle, which became the demo kit of the offer. An appointment-request agent takes a mail from the inbox, lets a model read it and puts the appointment into the calendar. The builder wrote the deployment as a service, a test procedure, a backup and restore, and an upgrade without downtime. The reference implementation was built afterwards and tested live with five test mails. A run costs about 0.003 dollars on a small model.

Sales sharpened the offer day by day. It collected three objections from the target group with sourced answers, then added one argument from the builder's progress each day. Content turned each day into a behind-the-scenes draft, every statement with a path to its evidence.

What went wrong

  • One day was lost on 28 August. The scheduler only ran while the desktop app was open. Triggers belong on a process that does not depend on the desktop.
  • The ops agent reported a file as missing twice although it existed, and it summed its delegation counter across days. Reports need a mechanical file check, not only the agent's memory.
  • On day one an agent started its own background process, which ended when the trigger ended. The trigger now demands that work happens in the same reply.
A rule file of one page produced five days without a single forbidden action.

Three things I took from it

Rules in plain text carry the system. One agent per area works better than one agent for everything, because each role built on the previous day and the handovers ran without me. And the report is the product for the founder: one file a day instead of four chats.

Outcome

The pilot became a fixed-price package: 4,900 euros net for setup, 490 euros a month for care. Founder time was not measured. The budget was three to five hours a week and I used only the time to read the reports.

Hung Mai

Hung Mai is an independent consultant in Weilmünster, Germany, mostly working remotely. He plans, builds and monitors AI automation for marketing and sales for companies without a digital team. About Hung Mai

Let's build something real.

Let's talkResponse within 24h.