Page 2 of 5
Thoughts on AI engineering, responsible automation, and building systems that matter.
Debugging an AI agent is not like debugging a normal function. The failure can live in the prompt, the retrieved context, the tool result, the memory layer, or the model's confident interpretation of all of it.
Read more →
The question I keep coming back to is not whether an AI agent will make mistakes. It will. The useful question is whether the system around it makes those mistakes visible, contained, and reversible.
Read more →
The third time a broken test made it to the remote repository because a developer forgot to run the test suite locally, I decided to solve the problem permanently.
Read more →
I used to start every morning with three browser tabs, three inboxes, and a growing sense of dread. Now an AI agent reads them all before I wake up. The journey from inbox chaos to automated calm took six months and changed how I think about attention.
Read more →
At my first startup, documentation meant a README with three bullet points and a prayer. Then I joined a German enterprise and discovered that documentation is not overhead - it is infrastructure. That shift changed how I build everything.
Read more →
A polished demo proves that an agent can succeed once. Production readiness means knowing how often it succeeds, how it fails, what it costs, and whether a person can safely recover the workflow.
Read more →
An AI agent that only says what it did is not enough. If it can act inside real systems, it also needs to explain why it acted, what evidence it used, what it ignored, and where a human can safely disagree.
Read more →
My first week at the new company, I tried to paste a code snippet into ChatGPT. My colleague physically reached over and closed my laptop. "We do not do that here." That was my introduction to real AI governance.
Read more →