Here's How 3 AI Agents Help Me Day to Day

Three AI agents run on my server and they mostly show up as small interruptions: a message when I wake up, a ping in the afternoon, a note that was not there yesterday. Here is what a day actually looks like.

Morning

The main agent handles the morning briefing. It reads my email, my calendar and yesterday’s notes from the second brain, merges it with the day plan, and sends one message to my phone. By the time I am awake the useful stuff is already separated from the noise. This is the generalist, the one I actually talk to during the day. Research, code, smart home, writing, it does the heavy lifting and keeps the whole context.

A few minutes later Hermes Health checks in. It pulls the overnight sensor data from Home Assistant, compares it with the habit checkboxes I ticked yesterday, writes a short insight into today’s daily note, and messages me the summary. Sleep, weight, watch battery. It is the one that connects three bad nights to a sliding chess rating. Nobody asked it to, it just does.

Afternoon

Mid-afternoon a scheduled job moves calendar events from my phone into Vikunja. I add the events during school hours from my smart watch, because phones are banned in school and the watch is not. By the afternoon the tests and deadlines sit where the study bot can see them.

Late afternoon is the study coach’s turn. It reads the next two weeks of tasks and checks the prep windows. If a unit test is seven days out, one message arrives: subject, date, what to review, minutes for today. If nothing is in a window, it stays quiet. The daily cap is one message, and it has never tried to guilt me into homework. I do not do homework. It knows.

Evening

Everything goes quiet on purpose. The health bot does not message after morning unless I ask. The study bot only comes back the night before a German vocab test. I wrote that into both of them myself, because a bot that nags all evening stops being helpful and starts being a notification badge.

Why the fleet works

Every agent is narrow by design. The health coach cannot touch my school schedule. The study coach does not care about my habits. Neither of them carries the generalist’s full memory, and that is exactly the point: smaller context, faster answers, fewer mistakes. Separate Telegram bots also feel like separate people, which sounds silly until you have talked to a dry practical study bot after a warm health bot.

Underneath they share everything. Same server, same Home Assistant, same Obsidian vault, same MCP servers. Three windows into one stack, each with a locked scope, no single bot owning the data.

The day runs itself. Two short messages, notes that write themselves, data flowing where it is useful. I read, I act, I do not manage logistics. That is the whole idea.