Early this month, my trusty development laptop broke. Thanks to shipping problems in Vietnam and severe weather in Kentucky, the replacement was delayed for ten days. I was stuck with the computer we use for videoconferences, without any of the tools we use for writing or testing software. How could I do my job with just a web browser?
Fortunately, I already had some experience with coding agents. You assign a task to these AI bots, let them work on it for a while, then review their work once they’re done and either accept it or request changes. Our early results had been mixed, but the technology had improved a lot since our experiments in 2025. What could I accomplish with the new tools like OpenAI Codex and Github Copilot?
GitHub lets you assign a software issue to an AI Agent instead of a human.
I wasn’t willing to let AI agents work on core FromThePage functionality, since my ability to test the resulting code was limited to the hours I could borrow Sara’s laptop. But we had some long-standing “technical debt” in our automated test framework–the tools that prevent new bugs from arising in old code whenever we build new features. Ten days later, the bots and I had revised this test code, improving hundreds of files and expanding the code it tests by as much as we usually accomplish in a year.
But the day to day process was draining. I worked much longer days at much lower intensity than I usually do. Rather than focusing on a single task for an hour, I was bouncing between four or five different agent sessions, advancing each one, then waiting fifteen minutes to repeat the cycle. Since the bots were always running, and since the intervention needed was short, I never really stopped working. Waiting for my morning coffee to brew? Let’s check the bots. Up in the middle of the night? Let’s check the bots.
When the agent's code breaks, another agent can try to resolve the problem.
This experience is called “bot-sitting”, and a report on the phenomenon by the Work AI Institute studies it in depth. They analyze reports from heavy AI users in the corporate world, finding that heavy (but uncritical!) AI use leads to fatigue as humans find themselves cut-and-pasting between different agents. They start to cut corners, passing along results they don’t really understand, stop interrogating the AI outputs altogether, and even eventually experience “moral disengagement”. According to their survey, “heavy AI users are 3.4x more likely than light users to blame the tool when something goes wrong”.
At FromThePage, we value meaningful work. I can’t stand the idea of reducing transcription to brainless button-clicking that only produces slop. And yet, we’ve embraced AI tools to help volunteers get a second opinion on difficult words or create first drafts. How can we keep users from experiencing this kind of fatigue and moral disengagement?
We have some ideas and are running a few experiments, but we want to learn more. I’d be interested to hear from people who have run into this problem. Have you experienced this fatigue, or seen it in your colleagues, students, or volunteers? What are you doing about it?