My personal agent setup

Why it keeps getting better

AI is, I would say, the first technology that can help improve itself. So a team that tells its AI the problem and has it fix its own way of working keeps getting better, while a lot of people try it once or twice, see it fail and go back to the old way. More of the gains come from designing the process than from the model alone.

My judgment and approvals sit on top. Below them, the written rules and memory every session starts from, methods that load when a task needs them, the choice of model, tools that reach my work, checks that can refuse an action, and the scheduler that wakes standing jobs. Around them all, a loop turns each new failure into a fix.

Where it began

After fourteen years in go-to-market and operations, I started working this way at SEOmonitor, which sells search analytics software to SEO agencies. Handing more of my own work to AI agents was the natural next step. It was driven by curiosity.

How it grew

I kept meeting limits. Some sat in the models, others in the harness, the software around the model that holds its rules and its memory. The rest were in my own processes.

What changed is how I work. When I name a problem, the agents research it and write the fix into the setup as a rule, a check, a tool or a method. Every later session starts with that fix.

That also meant finding where its own rules failed. The setup said that work which decides what gets built must start with a few questions to me, each with the agent's best guess for me to correct. When I measured it, most sessions had never asked me anything, and I had to correct their work afterwards. So a check now fires when a session is about to turn its plan into work without having asked me anything. It stops it there until it asks.

What the setup is made of

A new session can open on a project it has never seen. It reads the written plan and the project's board, and it names its next step.

When I ask for a status report, the agent loads the reporting method and follows its steps instead of improvising.

Most of the work runs on flat-fee plans from more than one AI company, and pay-per-use models only run inside a small set budget or with my approval. A search through hundreds of log files for one error goes to a simple script or the cheapest helper, while a strategy memo stays on the strongest model, set to think longer.

The tools let the agents read and act where my work happens. They reach my email, my calendar, my messages and the web. The web part matters, because a plain agent that tries to read a page often gets an error, an empty page or a page that doesn't load properly, and then it gives up without finishing the task. So my setup reads in steps. A normal read comes first, then other free methods if that fails, and a paid reading service only as the last step. 369 reads still worked this way after the first attempt failed. The agents read, draft and answer email.

The computer's scheduler wakes standing jobs at set times. One standing job is my technology radar. Each day it collects newly released tools for AI agents and sorts each into look, watch or skip, and a tool is adopted only after a test on real work.

Other standing jobs audit the setup itself and improve it. And the same rules and methods are mirrored for the other agent tools I use, Codex, OpenCode and Google's Antigravity, so none of them starts from zero.

One piece of work, step by step

If the work decides what gets built, an agent opens with a few questions to me, each with its best guess, so I correct them instead of writing answers. An agent then starts from the written plan and the project's board and names the next step, because recent chat alone forgets. Agents do the work, each with the method the task needs and the model that meets its quality bar, and smaller parts go to helper agents. Before anything risky runs, an automatic check can refuse it, so a known mistake is caught before it repeats. An agent that didn't do the work checks the result itself, never only the helper's report. Anything published, spending past a set amount or that can't be undone waits for my approval, on a short card with the agent's recommendation; I click an answer and the work goes on. A new kind of failure goes back into the setup.

Where a check can say no

Automatic checks run before an agent acts and can refuse when a known mistake is about to repeat. When an agent tries to email a person or company that already got an email from me in the last 14 days, the check refuses it, shows the earlier thread and tells it to reply there, and the agent does. A written rule on its own controls nothing, so the check sits on the only path to a risky action.

From directing to doing

I had run marketing programs as a manager and a director, never hands-on. With this setup I did the work myself, and the search program at Flip Education, the AI co-teacher for lesson planning that I co-founded, brought in 37,300 clicks from Google over ten weeks, from March to May 2026. The search program has its own page on this site, where I tell the whole story.

What I would tell the team that owns this work

Don't be afraid of experimenting. Look at the many models out there. And research what world-class looks like in your field, so you can judge what the AI gives you.