Productivity theatre, automated

Agent swarms, software factories, a dozen agents running in parallel. Every few weeks there is a new name for doing more with AI, and everyone is keen to show how much more they now get done.

A lot of it is productivity theatre. It looks like work, it burns tokens like work, and it ends up on a slide nobody reads. Most of us know someone with a new app, a new plan and several agents running all day, whose business still hasn't launched. Sometimes that someone is us.

Agents don't lower the bar for intent. They raise it. When starting something costs almost nothing, deciding what deserves your attention becomes the whole job. Four principles help.

1. Count the human work that comes after

We tend to value an AI result by how hard it would have been without AI. This report would have taken a week. This prototype would have needed a developer. Those comparisons say something about capability. They say nothing about whether the result deserves your time.

The agent's part is only the first cost. Then you have to read it, understand it, decide what you believe and work out whether anyone cares. Send it to your team and they all do the same. Make it a weekly report and you have created a weekly claim on several people's attention, set up in seconds.

So follow every result through: who will use this, what happens next, and is that work I actually want? A contractor who has a model read blueprints to prepare quotes learns quickly that it needs to show its measurements, because the measurements decide the price. A more polished quote wouldn't help; a checked one would. Once you know what a person has to do with the output, you can point the agent at that part: check the evidence behind a recommendation, find the assumption that would change the answer, resolve the question you would otherwise have to investigate yourself.

2. Explore more than you commit to

One of the best things agents give you is cheap exploration. You can chase a question you couldn't afford to before, try an approach and drop it, or find out that an idea you loved doesn't survive contact with reality. That is valuable even when you keep nothing.

The lightness disappears when every experiment turns into a project. Projects come with decisions, feedback, maintenance and a place in your week. Starting is now so cheap that you can take on more responsibilities in an afternoon than you can carry in a month.

Say customers try your service once and don't come back. You could send agents to look at it from different sides: one through the cancellation reasons, one through the help requests, one comparing what returning customers buy with what the others bought. Different explanations may come back, and each points to a different decision. Or you could ask one agent for a loyalty programme, another for a new pricing page and a third for win-back emails. They may all look excellent, and now you are reviewing three solutions before you understand the problem, each with its own next steps and the pull of work already done.

Parallel agents help most when each branch can teach you something the others can't. Five agents producing five versions of the same answer just hands you a comparison job. And sometimes the missing piece isn't something an agent can produce at all. If the question is whether customers will pay, someone eventually has to ask them to pay. Another simulation feels safe. It is usually avoidance.

3. Decide when it stops before you make it recurring

A recurring agent can easily outlive the reason you created it. You asked for a weekly update on a market while you were considering it. Two months ago you chose a different market. The update still arrives, accurate and unread, spending tokens every week. That isn't the agent's fault. Nothing told it the decision was made.

It is so easy to turn a moment of curiosity into a recurring job that it feels clever every time. Each one is small. Together they shape your day around questions you no longer need to answer. Give every recurring job a stop condition up front: run for three days, stop when a pass teaches nothing new, stop when the decision is made. If you want it again, start it again.

The same goes for curiosity. Let agents surface the unexpected, but you decide what is worth following. You don't have to chase a thread because an agent found it interesting.

4. Give your coordinating agent an attention budget

Many people now run a coordinating agent, a kind of chief of staff that keeps the others busy and delivers a neatly organised set of demands every morning. Organised demands are still demands. Agents have near-infinite attention; you have a day, a life and a limit.

The useful version knows your commitments and what a new assignment would add to them. Research into a new market is cheap for an agent and expensive for you, if it leads anywhere. Brief it on what deserves an interruption: a decision only you can make, a commitment about to slip, new evidence that would change something you have already decided, perhaps the kind of customer story you always want to hear. Everything else waits. Its job is to prepare choices, explain what each would require, and help carry out the work you approved, so your attention can go somewhere else.

Theatre or work?

Three questions are worth asking about every agent you are running right now. Which ones help with something you actually care about? Which are experiments you are still learning from? And which are running only because you forgot to tell them to stop?

If those are hard to answer, some of what looks like productivity is probably theatre. It helps to ask your agents for the same honesty: not just "this matters, decide now", but "this is only an experiment" and "this probably doesn't matter". You have more output than ever. What is scarce is intent, and that part is still yours.