Blog · 12 July 2026
The AI hour is turning into a review hour. Your tracker misses the quiet part.
AI writes more of the code now, so you spend more of your day pointing it in the right direction and checking what it produced. That is real work. But when you are reading a running agent's output and steering its next move, your keyboard often goes quiet, and a tracker that decides you are working by watching for keystrokes reads that quiet as idle. Here is what the research shows about how big the review-and-supervision slice has become, and how to count the part of it that every other tracker throws away.
The balance of AI work is shifting toward review
Google's DORA program runs one of the largest ongoing studies of how software teams actually work. Its2025 report finds developers now spend a median of about two hours a day on AI-assisted work, close to a quarter of the working day. And it is direct about what that does to review: as AI writes more code, reviewing and supervising that code becomes more important, not less.
GitHub's Octoverse 2025 shows where the extra review load comes from. AI agents now open pull requests at a scale that did not exist a year earlier: GitHub's own coding agent opened more than a million pull requests in five months. When a machine can produce review-ready changes while you do something else, more of your day becomes directing that work and deciding whether it is right.
None of this means writing code has gone away. On DORA's own numbers, writing is still the single most common thing developers use AI for. The point is narrower and more useful: the mix is tilting, and a growing share of the day now goes to supervising and reviewing rather than typing.
The quiet part is the part that disappears
Reviewing AI's work is really two different activities. Part of it is hands-on: you are reading a diff in your editor, typing a comment, scrolling a pull request. That time is keyboard-active, and every tracker already counts it as ordinary work.
The other part is different. You start an agent, and for the next few minutes you are watching it run, reading what it writes, and deciding where to point it next, hands off the keyboard. You are working the whole time. But there is no input to detect, so an activity tracker marks you idle and drops the minutes. A manual timer does not save you either, because in the middle of directing an agent you do not stop to start a stopwatch. This is the quiet part of the review hour, and it is exactly the part that goes uncounted. (We put a dollar figure on it in a separate post.)
The fix: a third state
Counting all of your idle time would overcharge, because some of it is genuinely stepping away. The trick is to tell two kinds of quiet apart: attending a running agent, which is work, versus being away, which is not. Atend adds that third state. It notices when a known agent is running and whether you are there, counts the attending time, and leaves out the time you actually stepped out.
It does this from metadata only: whether an agent is running and whether you are present, never your code, never your screen, never a keystroke. So the quiet part of your review hour finally lands on the clock, and Atend never reads a line of what you wrote.
Atend automatically counts the time you spend attending an AI agent, no keystrokes, nothing to log.
See your own number
Atend reconstructs your real day automatically, including the quiet attending hours other trackers drop.