Managing AI agents when you work alone

Managing AI agents sounds like a job for a department. When you work alone, it is a job for you, on a Tuesday, between other things. The skills are the same ones a good manager uses with people: say clearly what you want, agree on what finished means, and make it easy for them to tell you when they are stuck. The difference is that an agent never gets tired of writing things down, so you can ask for more of that than you would of a person.

Hand over work one ticket at a time

The most common mistake is handing over a wish instead of a task. "Sort out the website" is a wish. "Replace the old opening hours on the contact page" is a task. An agent will happily start on a wish, and you will get something, but probably not what you meant.

So put each piece of work in its own ticket:

• A title on one line. What needs to happen, not a story about it.
• A few lines of detail. Where to look, what you already know, what not to touch.
• What done looks like. One or two sentences you could check in a minute.

If a ticket needs three paragraphs of context, it is probably two or three tickets. The guide on project vs task helps with where to cut.

Statuses are the shared language

You and the agent need a way to know the state of the work without a meeting. A small set of statuses does that. Something like this is enough:

• New. Filed, not started.
• In progress. Someone is working on it right now.
• Blocked. Waiting for you, with a question.
• Done. The work is finished and described.
• Deployed. The result is confirmed live.
• Cancelled. It will not be done, and the thread says why.

Ask the agent to set the status every time the state changes. Then a glance at the project tells you what is moving, what is waiting for you, and what is finished. That is most of managing AI agents, right there.

Blocked means a concrete question

An agent that is stuck will often keep going and guess. Tell it not to. When it needs a decision, it should set the ticket to blocked and write one question you can answer in a line.

"I need more information" is not a question. "The two price lists disagree on the delivery fee. Use the newer one?" is. Blocked tickets are the first thing you read when you open the project, so make them quick to answer.

Review done before it counts

Done means the agent says it is finished. That is a claim, not a fact yet. Read what it wrote, look at the result, and only then treat the work as finished. If it is not right, reply in the thread and move the ticket back to in progress. No drama, no rewriting the ticket from scratch.

Over time you learn which kinds of work the agent gets right on the first try. Review those lightly. Review new kinds of work carefully.

Done is not deployed

Keep finished and live apart. A draft that is ready is done. A page that is published and checked is deployed. An email that is written is done. An email that is sent is deployed. The gap between the two is where you decide.

This matters more with an agent than with yourself, because an agent can make "finished" sound very convincing. Asking it to confirm that a change is actually live before marking it deployed saves you from discovering later that it never went out.

Read the thread instead of re-asking

Every ticket has a thread: what the agent tried, what it found, what it changed, what it asked, what you answered. Read it before you ask "what happened with this?". The answer is usually already there.

The thread also helps the agent. When it picks up related work next week, it can read what was decided last time instead of starting from nothing.

One agent per project

If you use more than one agent, give each its own project, or at least its own clearly separate work. Two agents editing the same thing at the same time create the same mess two people would, only faster. One agent per project also keeps access small and makes it obvious who did what.

How often to check in

You do not need to watch. A rhythm is enough:

• When something is blocked. Answer soon. A blocked agent is an idle agent.
• Once or twice a day on active projects. Skim what moved to done, review what matters.
• Once a week across everything. Close, cancel and re-prioritise in a weekly review.

If you find yourself checking every ten minutes, the tickets are too vague or the agent is being asked to decide things that should be yours.

It is still your project

Managing AI agents well is mostly good personal project management with a helper added. You decide what the work is, what done means and what goes live. The agent does the defined work and reports. For the wider setup, see using an AI agent for project management.

Very Simple Projects is a simple personal project management app built for this: projects, numbered tickets, a thread per ticket, and six statuses plus your own. You connect an agent to one project, it files tickets, reports in the thread, sets blocked with a question when it needs you, and marks deployed only when the change is confirmed live. It costs $0.99 a month, and the first month is free.