AI agents, built with rules and handed over
An AI agent is only as good as the rules it follows. I write those rules with your team, build the agent, and hand it over so you can change it.
What I mean by an agent
An agent is AI that does a job on its own schedule instead of waiting for someone to type into a chat window. It checks for new work, does what the process says, and reports back.
That's powerful, and it's exactly why it needs limits. An agent that guesses wrong does it quietly, again and again, until somebody notices.
An agent I built
For a software startup, I built an agent that runs the sales pipeline three times a day. Each run, it picks up new signups, matches replies to the right person, spots bookings and no-shows, turns call transcripts into notes and tagged product feedback, and drafts the follow-ups that are due.
It works inside four rules:
- Drafts only. It never sends an email. The founder reads and sends every one.
- Never twice. It never emails the same person twice about the same thing.
- Ask, don't guess. When it isn't sure, it asks in its report instead of guessing.
- Always a reason. It never marks someone as lost without writing down why.
After each run, the founder gets a short report: what came in, what was matched, what was drafted, and what it wasn't sure about. People keep using what they can see.
The full reasoning is in Drafts, never sends: keeping a human in the loop.
How I build an agent
- Define the job. The trigger, the inputs, the steps, the judgement calls, and what it must never do.
- Write the rules into the process. Not "please be careful" in a prompt. Where the tools allow it, the agent only gets the access it needs, like permission to create drafts and nothing else.
- Run it strict, on real work. A person checks everything at first. Rules loosen only after the output has been good for a while.
- Hand it over. The rules are in plain language so the person running it can change them. The founder at the startup set the whole thing up herself from one pasted-in prompt.
Which AI
I build agents on whatever fits the job and what you already pay for: ChatGPT, Claude, Gemini, Copilot or others.
Common questions
What does an AI agent consultant do?
Defines the job an agent will do, writes the rules it must follow, builds it on your tools, and hands it to your team to run and change.
Will an AI agent act without a person checking?
Only where a mistake is cheap and easy to undo, like an internal reminder. Anything that reaches a customer starts with a person approving it.
How often can an agent run?
As often as the work needs. The startup's pipeline agent runs three times a day.
Can a non-technical person run an agent?
Yes. The startup's founder doesn't code. She runs the agent herself and can read and change every rule.