How I run several businesses with a small team of agents

I am not an engineer. I have never shipped production code. I run food and beverage businesses, and for a long time the ceiling on all of them was the same thing: me.

Over the past months I rebuilt how my day works around a small set of agents. This is what actually happened — the parts that worked, the parts that failed, and what I would tell another non-technical founder starting now.

Where I started: the bottleneck audit

I spent one week writing down every task that only I could do, and how long each took. The list was unflattering. Roughly two thirds of my week was not judgement. It was transfer: reading an inbound message, understanding it, converting it into a quote, a follow-up, a document or an update somewhere else.

That is the work worth handing to an agent. Not strategy. Transfer.

Communication

The first agent I built does nothing clever. It reads inbound messages, classifies them (new enquiry, existing order, supplier, noise), pulls the relevant context, and drafts a reply in my voice.

It does not send anything. I approve every message that leaves.

The gain was not the writing. It was that I stopped context-switching. A morning that used to be forty fragmented decisions became one review session.

Sales

For enquiries, the agent produces a structured brief before I ever reply: who they are, what they appear to need, what we have quoted similar buyers, what is currently available, and the two or three questions still missing.

The effect on speed is the real outcome. In trade, the supplier who answers first with a coherent offer wins a disproportionate share of business — not because they are cheaper, but because the buyer can move on.

Pricing

This is the one I was most nervous about, and the one with the clearest return. Market inputs get collected daily, normalised, and compared to what we quoted last. When something moves beyond a threshold, I get a note telling me what changed and which open quotes are affected.

I still set every price. The agent removed the excuse of not knowing.

What I deliberately did not automate

• Anything that commits money or makes a promise to a customer without me reading it.
• Relationships. A supplier call at harvest time is not a workflow.
• Judgement calls where being wrong is expensive and slow to detect.

What I got wrong

I built too much at once. The first version tried to handle four processes and I could not tell which part was failing. Rebuilding it one workflow at a time, each with its own log and its own owner, was slower and worked.

I also over-trusted early. An agent that is right 90% of the time feels excellent for two weeks and then costs you a customer. Review gates are not training wheels; they are the design.

Where this is going

I am still building. The next layer is agents that talk to each other with clear handoffs, so a lead becomes a brief becomes a quote becomes a document with me approving at three points instead of forty.

The reason I write this down is that I advise organisations on exactly this, and I would not trust advice from someone who had not run it inside their own business first. Nothing here came from a slide. It came from a week where I was the bottleneck and decided to stop being one.