Contact

What I actually want to hear about

How you work. Not the tool list — the practice. How you get code, docs and comments out of an AI session that a human can still read and reason about months later, when the session that wrote them is long gone and nobody remembers the context.

That is the hard part and the least written about. Most of what circulates is about getting the model to produce something that runs. Very little is about getting it to produce something a person can pick up and understand.

So: what does your loop look like? What did you change after something went wrong? What did you try that did not survive contact with a real project? An incident and the thing you changed because of it is exactly the shape every note here takes, and I would rather read yours than write another of my own.

Also worth sending: a finding here that did not reproduce for you. Everything was observed on specific models in one harness, and this repo’s own argument is that findings are model-relative — a result that fails on your stack is data, not a complaint.

How to reach me

• @mivanov_dev — on X. The quickest way to get a reply, and the right place for anything short.
• mivanov.dev — my site: other writing and projects. It is a live project rather than a finished one, so it may be down when you try it. X is the reliable route.
• Open an issue — for anything that should end up in the repo rather than in a thread.

The repo

Everything here is markdown in one repository, MIT licensed. Read it on the site, or take the whole thing:

git clone https://github.com/mivanov93/ai-agent-field-notes.git

The source and the issue tracker are at github.com/mivanov93/ai-agent-field-notes.

One use worth knowing about: point a session at the whole repo and ask it to critique how you work. The notes are written to be read that way — short, concrete, and each ending in something a model can act on without you translating it first.

Corrections and prior art

The standing rule: “I have not found prior art” does not mean “first”. The searches behind these notes were a handful of research agents over public 2024–2026 material — a real method with real gaps. Show me earlier work for a claim and it becomes a citation.

That applies to me too. Bans rotate the vocabulary partially corrects vocabulary control, because I measured my own repo and half of my original claim did not survive.

I am not checking every day, but anything that arrives gets addressed in due time.