AI Did Not Expand Your Attention

Something changed last December, I went from constantly correcting and steering coding agents to looking at what Opus produced and thinking, damn, I would have needed three passes over a week to get there.

When I write code, I usually arrive at the final version through iteration. I try an approach, find the awkward edges, restructure it, run the tests, and look at it again. Eventually, I mold it into something that feels elegant. Opus jumped straight to elegance within five minutes (yeah, yeah, yeah, I had good instructions in Markdown files, but still).

There is an illusion in this. The code may look finished, but producing code is not the same as finishing work. I still need to understand it, review it, integrate it, and determine whether it solves the right problem. The compression, however, is real enough that it changes how you think about your capacity.

If one useful result takes two minutes, it feels reasonable to ask for ten. An agent can investigate one direction while another implements something else. A third can review the implementation, and a fourth can research the problem you plan to work on next. Suddenly, it appears possible to work on everything at once.

It is not.

Output Scales. Attention Does Not.

AI has dramatically increased our capacity to generate work. It has not increased our capacity to care about that work. Every result still needs attention, judgment, and someone willing to remain responsible for it.

In practice, I can give proper attention to one or two active agent sessions. One of those sessions might be a lead agent coordinating several others to research a bug, plan a fix, and implement it. From my perspective, it stays one attention surface. I can follow the reasoning, drill into the work, and remain responsible for the result.

I have experimented with more and find that having more than three agents doing unrelated work means I’m no longer the director, just a guy monitoring queues. Small, focused teams outperform large diffuse ones for the same reason. The coordination cost grows faster than the output. The same constraint applies to agent sessions. More is not the direction to optimize in.

Saved Time Is Not New Time

I’m a huge fan of the FranklinCovey framework based on “making the main thing the main thing,” and I say that as someone who has built engineering teams on it. The core insight is about understanding our relationship to time vs. tasks. We talk about having more time, saving time, and managing time, but none of those things is quite accurate. We have all the time there is, we have to spend it every day, and the only thing we can manage is ourselves within it.

AI can save time on a task, but it cannot give me more time. I still have the same week. I still have to decide what that week is for.

That distinction is easy to lose when an agent completes in minutes what previously took days. The saved time feels like newly created capacity, so we immediately spend it. We start another task, then another. Each one produces code to review, decisions to make, and new directions to consider.

This is how someone can work 12-hour days with AI and still ship very little that matters. The agents remain productive and the person remains busy, but the goal barely moves. Increased execution capacity gets mistaken for increased capacity to direct work.

Covey’s Time Matrix is useful here because it separates urgency from importance. Responding to every announcement, trying every new tool, and rebuilding a workflow after every model release feels urgent, but is usually busywork. Its urgency is manufactured.

The second quadrant is where the useful AI work mostly belongs: choosing where it creates leverage, staying responsible for the output, improving the systems around it, and deciding what should be built before asking an agent to build it.

The time AI saves should create more room for that work. Instead, we often return it immediately to urgency, interruptions, and more generated activity.

Low Friction Has a Cost

AI also removes a useful constraint. Before agents, trying another direction had a real cost. You had to decide whether an idea was worth several hours or days. That friction forced you to choose.

Now another direction costs a prompt. This is useful when exploration is the goal. I can have an agent try ten approaches, but I need to be clear that the goal is exploration. I time-box it, evaluate the results, pick a direction, and stop. Without that boundary, cheap exploration becomes expensive distraction.

Experienced project managers schedule to roughly 65% capacity. The rest absorbs reality: the interrupt, the misread requirement, the test that reveals a wrong assumption. AI does not change this math. It changes what fits in 65%, not whether the buffer is needed. If anything, the buffer matters more now, because AI can fill the remaining 35% with convincing-looking work that still needs your judgment to become real.

Point Your Agents at the Main Thing

So how do you tackle this? Start with your goals, OKRs, or whatever you call them. Then plan weekly, and finally daily. The order matters even more when agents can generate work faster than you can evaluate it. Your goals determine the week’s priorities, and those priorities determine what you ask the agents to do.

Reverse that order and the agents generate the plan for you. A result exposes a possibility, the possibility becomes a prompt, the prompt creates another result to review, and the agenda sets itself. The tool stops being a servant, and technology is a great servant but a horrible master.

Daily planning is not enough to prevent this, because the day is already too reactive. The constraint has to be set earlier. The goal for the week has to be established, the few things that would meaningfully advance it have to be named, and then the agent sessions have to be pointed at those things specifically.