Screen time doesn't measure attention

A writer measuring his own use of the Hermes Agent, run locally on a Raspberry Pi, found that screen time undercounted his actual interaction with the AI agent by a wide margin: Telegram logged only 6.6 hours that week, while his calculated interaction time — measured from first message to the agent's last reply within blocks separated by pauses of 40 minutes or more — totaled a full day per week. Of the pauses in his usage data, 88% were shorter than 20 minutes and 9% were longer than an hour, with only 3% falling between. The writer said the agent was meant to let him be more present in his life while it did the work, but instead it "is constantly pulling my attention into its orbit," and he is considering switching from Telegram to a more asynchronous interface such as email.

My girlfriend complained that I wasn't really present during our conversations and was hard to talk to. This happened during a week of excessive vibe coding on my phone. I know that sounds funky, but I use Telegram as my messenger to chat with my AI agent. I knew she was right, and I could feel it too. My head was constantly buzzing, and I didn't sleep well. As if on reflex, I checked my screen time. Given how I was feeling, I expected to see some big numbers for Telegram, but I didn't. I had spent "only" 6.6 hours on it that week. Unconvinced, I took a closer look at how I actually used my agent.<sup 1 footnote-1 </sup My theory was that I interact with it in blocks. For example, I give it a task, then do the dishes, and when it reports back, I give it another task or further instructions. I don't have Telegram open all the time, but my attention keeps orbiting around it. To measure the length of these blocks, I had to identify them first. That meant finding the actual breaks in my usage data. Of all the pauses, 88% were shorter than 20 minutes, and 9% were longer than an hour. Only 3% fell between the two. Knowing myself, I'd say an hour is enough to detach, while 20 minutes is too short. The sweet spot had to be somewhere in the middle, so I just picked 40 minutes. That finally let me calculate my interaction time: the time from my first message to the agent's last reply within an interaction block. The chart below compares it with my screen time. Looking at it hurts a little. As you can see, I actually spent a whole day a week interacting with my agent, so my screen time was a massive underestimate. That wasn't what I had in mind when I set this up. My vision was that the agent would allow me to be more present in my life while it did the work. Instead, it's constantly pulling my attention into its orbit. I find it fascinating and obvious how far apart screen time and interaction time are. Obvious, because anyone who experienced a honeymoon phase knows how much a chat can occupy your mind even when your phone is in your pocket. Fascinating, because I've always equated screen time with the attention I give an app. As it turns out, that's not always the case. Either way, I need to fix this, and looking at the right data is a first step. I'm also wondering whether Telegram is actually a good interface for communicating with an AI agent. It does promote synchronous interaction by showing you when the agent is typing. That makes the agent seem more human, which I honestly like. But at the same time, maybe this environment is nudging me into interacting with it more and more. Other interfaces, like email, lean more toward asynchronous communication. I'm thinking about giving that a try. I use the Hermes Agent locally on a Raspberry Pi. It keeps a local database of every session, and that's where the numbers in this post come from. I dug through the same database in an earlier post https://heyjonny.dev/posts/i-think-i-used-the-hermes-agent-wrong/ . ↩ footnote-ref-1