The Day I Realized My AI Was Talking Too Much
I have a problem. And I suspect I’m not the only one. My Hermes AI agent — the system I’ve been building for months — sends me about 25 to 40 Telegram messages every single day. Cron job reports, AI news summaries, backup statuses, battery alerts, genealogy updates, vault posts, system health checks. All useful. All well-intentioned. But together, they create a constant stream of noise that I stopped actually reading weeks ago.
Last week, I sat down and looked at my Telegram history. What I saw surprised me: I was scrolling past most of my own AI’s messages. The irony wasn’t lost on me. I’d built this elaborate system to help me stay informed, and instead I’d built a notification machine that I was training myself to ignore.
The Numbers Didn’t Lie
I mapped out the exact message flow. The main cron output channel (topic 4493) was pumping out ~15 messages a day. Agent interactions (Sara, Alice, Jeeves, Fred, Bob) added another 5-10. My direct chats with Charly, my system reflection agent, contributed 5-15 more. And scattered across various research topics, a few more sporadic messages. The total: 25-40 messages. Every single day.
Now, I’d already done some optimization. The morning briefing cron bundles the overnight reports into one message. The AI news goes directly to WordPress. Battery alerts only fire below 10%. And I’d implemented the [SILENT] protocol for cron jobs with nothing new to report. But it wasn’t enough. I was still drowning.
Three Solutions, One Insight
I analyzed the problem and came up with three possible solutions, each with different levels of effort and impact:
1. The Simple Fix — Silent Cron, Smarter Briefing
This is the low-hanging fruit. About 1-2 hours of work. The idea is simple: flip every remaining cron job from “deliver to Telegram” to “deliver locally” (store the output, don’t send it). Then add all those silent outputs to the morning briefing, which already bundles reports. Add an afternoon wrap-up at 17:00 for the rest. Result: 15 messages a day in the main channel drops to just 2 — the morning briefing and the afternoon wrap-up. That’s an 80% reduction for minimal effort.
2. The Dashboard — Mission Control Berichten-Dashboard
A more ambitious approach: build a message dashboard on my existing Mission Control website. A place where ALL cron output, reports, and alerts live — organized by category, searchable, filterable. John can log in and see what happened whenever he wants, without Telegram being the bottleneck. Critical alerts still go to Telegram. Everything else lives in the dashboard. This is a 1-2 day project.
3. The Overengineering — AI Post Filter Agent
And then there’s the geeky option: a dedicated Hermes agent profile (“Post”) that acts as an intelligent filter. Every outgoing message passes through it. It categorizes, prioritizes, extracts action items, and decides what reaches Telegram, what goes to the dashboard, and what gets silently archived. It even creates Google Tasks from detected action items (“battery at 0% → replace sensor”). This would take 3-5 days and cost about $0.23/month in LLM tokens. But honestly? It’s probably overkill for now.
What I Learned
The real insight wasn’t in the technical solutions. It was this: every message you send is a promise you’re asking someone to read. When you’re building a personal AI system, it’s easy to think “more information is better.” But more information is only better if it’s actually consumed. If your own system trains you to ignore its messages, you’ve created a negative-feedback loop that undermines the whole point of having an AI assistant in the first place.
I’m starting with solution 1 this week. Silent cron, smarter briefing. If that’s not enough, I’ll build the dashboard. And only if both fail will I reach for the AI filter. The lesson: sometimes the best optimization isn’t a smarter algorithm — it’s just sending fewer messages.
How many notifications does your smart home or AI system send you per day? Have you ever caught yourself scrolling past your own system’s messages? I’d love to hear your experience in the comments.
Wat vond je van dit bericht?
