How I run my own work with agents
The method
I run my own follow-up work on a small set of AI agents. They read my email, calendar, meeting notes and messages, keep track of what was promised and by whom, and prepare drafts. They never send or delete anything, and never change anything already in my mailbox or in our company system. A separate agent checks the whole system every day. Here is how it is built, what went wrong, and what I would tell any leader who wants to do the same.
The problem I wanted to solve
My work runs across many channels at once: email, meetings, calendar invitations, messages, documents. Something promised in a meeting is easy to lose by the time the email thread moves on. I wanted a system where nothing I promised, and nothing promised to me, could quietly disappear.
What the agents read
Every few hours, the agents go through my email, my calendar, my meeting notes and my documents. They also read my messaging conversations when I have them open on my computer. The list of channels is kept open on purpose, so a new channel can be added later without rebuilding anything.
They look at my own actions too, not only what others send me. If I already replied to something, or forgot to, the system should know.
What they do with it
- They keep a running list of commitments: what I promised and to whom, what others promised me, and what is still waiting.
- They link every item to its place in our company system: the client, the project or the request it belongs to. If an item has no obvious place, they ask me where it belongs.
- They prepare drafts. After every meeting I send a follow-up email, so a draft follow-up is prepared for each meeting.
- They follow every draft to the end: sent as it was, sent after my edits, deleted, or still waiting. If I delete a draft, the system asks me about it until I act or give a reason.
What they never do
They never send anything, never delete anything, and never change anything already in my mailbox or in our company system. They prepare; I carry it out. This one line is what lets me trust the rest. The page When should an AI agent stop and ask a person? explains how to draw that line.
The checker
A separate agent, an auditor, goes over the whole system once a day. It does not do the work. It looks for what went wrong: a job that did not run, an item that was lost between one run and the next, a report that does not match what is actually in the mailbox. What it is sure of within the system’s own working files, it fixes, and it writes down what it changed. Anything that needs my judgement, it brings to me as a plain question.
What went wrong, as it happened
In the first days, the system reported “nothing new” while drafts sat untouched in my mailbox. It also replied to me in English, when every message to me must be in Arabic. Each run looked correct from the inside. The faults were only visible from the outside, across many runs. That is why the auditor exists. The question Who checks the AI agent? tells this part in full.
What I watch most closely
The biggest risk I see is overload. The more the agents carry in their working memory at once, the less accurate they become. So decisions are written down outside the conversation, and each run reads only what it needs. The page What is a context window, and why does my AI agent forget? explains why this matters.
How I deal with it
I never open the system’s files myself. I deal with it only by talking to it, and I give most of my instructions by voice. When I decide something, the system records the decision, so the next run starts from it rather than from zero.
What I would tell another leader
- Start with one job you do every day, not a programme for the whole company.
- Write down what the agent must never do before you write down what it should do.
- Add a checker early. An agent left to grade its own work will tell you everything is fine.
- Keep agreements outside the conversation, in a written record the agents read again each time.