Monitoring & Observability

I have been working with monitoring and observability for more than ten years. It started in 2014 with homegrown solutions built on the ELK Stack. Today Grafana is at the center. It is the shared interface for me and for the AI agents that run my websites. This page describes how my view of monitoring has changed and why observability matters more than ever when agents are doing the operating.

2014: Homegrown log aggregation and profiling with ELK

Back when Elasticsearch, Logstash and Kibana were still not widely used, I built my own tools with exactly this trio. Two articles from that time are still published in the archive. One covers a homegrown Java profiler that sent measurement data through Logstash into Elasticsearch and displayed it in Kibana. The other covers a “homebrew edition” alternative to Splunk for central log aggregation.

The appeal back then was using free software to rebuild something that otherwise required expensive commercial products. Collecting logs from different sources in one place, then searching and visualizing them, was the core of what we now call observability, even though the term was barely used at the time.

Why ELK sits in “Assess” for me today

That is why the ELK Stack sits in the Assess ring on my technology radar. I keep an eye on it, but I don’t actively recommend it, and I no longer use it in ongoing project operations. I would do much of what I built in 2014 differently today. Instead of assembling my own infrastructure for collecting and analyzing logs, I would use cloud-native services that already handle operations and scaling.

The experience is still valuable. If you have built a pipeline yourself, from the log line all the way to the dashboard, you know where data gets lost and where latency builds up. You also know that a nice dashboard tells you nothing about the quality of the data behind it. That understanding shapes how I do monitoring today.

Today: AI agents run my websites

On bernd-zuther.de and ketovavaria.de, AI agents handle much of the ongoing work. They write content, deploy changes, check results and fix errors. That saves a huge amount of time, but it raises a new question: how do I know the work is actually done?

When an agent reports “done”, that doesn’t mean it is right. It only sees the world through the tools it has been given. If a tool is missing or returns incomplete information, the agent can confidently report success when the task has failed. So trusting agents blindly is not an option. A silent failure of a notification channel showed this in practice.

Grafana as a shared place for humans and agents

That is why I run my own monitoring, with Grafana as the central interface for both websites. For me, Grafana is in the Adopt ring, which means I use it actively every day. Here, observability closes a gap in two directions:

First, for me: I can see what is really happening on the websites without checking every single message from an agent. I look at data instead of taking reports on trust.

Second, for the agents: they can query the same data and check their own work against reality. An agent that has deployed a change can look at the monitoring to see whether the change actually has the expected effect, not just whether its command finished without an error.

This makes Grafana the shared place where humans and agents look at the same facts. That changes how we work together: discussions are no longer about who is right but about what the data shows.

From homegrown experiment to operational foundation

Between the 2014 ELK experiment and today’s Grafana setup, the direction is clear. Back then the goal was to get any visibility at all: collect logs, search them, analyze them, often out of technical curiosity. Today observability is an operational foundation. It is what allows me to hand responsibility to autonomous agents without losing control.

The basic idea is still the same: bring together data from different sources and present it clearly. What has changed is who it is for and what it is for. The dashboard used to be a tool for developers hunting down a problem. Today it is a shared source of truth for a team of humans and machines.

Outlook

My monitoring setup grows with every new task the agents take on. I’m planning a separate article on how it is built and what I’ve learned using it day to day. Until then, the main point is this: if you use AI agents in production, you need observability. It isn’t there to control them out of distrust. It gives humans and agents a shared set of facts.