Your AI Agent is only as good as what it learns
My day job is being an Engineering-Manager-slash-Tech-Lead. I started that in July 2025. The first half year my job was mostly managing humans and working on technical concepts and pitches. I had little time left for working on code. I started using Cursor to work on code and had Cursor rules in place (to instruct the AI on how to work with code, whata rules to follow, what not do do etc.) This worked ok.
I felt like I could start to focus on contributing technical work again.
It really changed once I was able to use Claude Code at work.
In this article I want to share what works best for me, and how the system I use to work on our products using Claude Code get smarter every day.
It all started, when I read this article from Will Larson Learning from Every’s Compound Engineering. I set out to create a reinforcement learning system for my agentic work. Every plan results in work done, which results in learnings captured. Those learnings then get used to improve the plans for the next work item.
I bet there is stuff to improve on here, but the results I get are good.
Merge Requests / Pull Request
Yesterday, while listening to Avdi Grimm’s podcast I thought about all the knowledge lost not transferred from code reviewers to me the AI. We do have thorough code reviews and sometimes errors, or architectural decisions impact the merge request significantly. Those discussions are only part of the MR and if you are not part of the discussion, you might not see it. The AI certainly does not.
That’s why I created a new skill, that combs through the last x merge requests and finds every relevant discussion by humans (and also by the AI review bot we have). It extracts the learnings and writes them to the project, so my compounding workflow can pick it up next time.
That’s it. Short and sweet. If you use AI agents to work on your code projects, make sure they get “smarter” every time.
Currently reading: Hum by Helen Phillips 📚 I am in part 2 of the book already. Hasn’t gripped me yet but it’s still a nice progression of the story.
Heiko bloggt regelmäßig (danke!). In dieser Ausgabe sind soviel gute Links dabei, dass ich einfach alles empfehle!
Today I successfully took my theoretical exam for the German C1 level driving license. Can’t wait to be able to drive that big motorhome.
This is not what I wanted to write. I wanted to write about how I’m about to go on book tour for my new book in a few days. Instead I am writing about the fact that I was just informed that my first book Let’s Pretend This Never Happened was banned
it is amazing what is unlocked by using Claude code for me. I build my own AI powered trainer and nutrition coach that helps me design training plans for the bike and I got rid of freshRSS and could build my own feed aggregator that uses the Google reader API, and can plug into all the reading apps.
Getting to know Current RSS reader by @tg@indieweb.social but have no idea what the difference is between releasing and article and marking it as read. And the docs also don’t explain it at all. Confusing!
I really values Simon’s (@simon @simonwillison.net) writing and the many insights I got regarding LLMs and AI in general. I would value even more if he moved his newsletter off of Substack. leavesubstack.com
Ich war heute mit den Kids beim Arzt. Kind 2 hatte ihr Eseltier mitgenommen. Um die Zeit zu vertreiben fanden wir alle Möglichkeiten, das Jäckchen möglichst falsch anzuziehen. Das führte dazu, dass Kind 2 sich vor Lachen kaum noch beherrschen konnte und einen hochroten Kopf bekam. Außerdem fühlte sich eine weitere Mitwartende motiviert, weitere Anzugideen vorzuschlagen. Wir hatten viel Spaß. Üblicherweise versteckt sich jede im Wartebereich hinter ihrem Handy. Etwas entspanntes Miteinander und gemeinsames Lachen tut einfach gut.
Zum Abendessen habe ich Reste eines Hokkaidos heißluftfrittiert, püriert und mit Milch und Frischkäse zu einer fein-sämigen Sauce bereitet. Diese gabs dann zu den Spaghetti. Simpel aber verdammt köstlich.
Neue Tubeless für BAB25 vom Steppenwolf
Ich habe mir vor über zwei Jahren ein Gravelbike gekauft und das seitdem über 1.100km gefahren. Nun steht am 11.10.25 ein Gravel Event, organisiert vom Steppenwolf an: Berlin—Angermünde—Berlin, 200km durch Brandenburg.
Weil es über sehr viele Kopfsteinpflaster gehen wird, will ich weiterhin tubeless fahren. Aber die Schlappen mussten mal neu gemacht werden, und das hab ich hier in dem kurzen Video mal verwurstelt.
(auch auf Insta www.instagram.com/reel/DO9E…)