Recent content on Kagamino's Blog
Coding agents have agency - they can decide on their own
what files they need to read in order to understand a codebase.
Common harnesses such as Claude Code or Codex provide a
limited set of tools to perform Software Engineering tasks -
Read, Write, Edit.
I read many news about LLMs nowadays. One topic I’m most interested in
is local inference for LLMs, and trying to find the sweet spot where they can be used
in local setups.
As I recently was sharing my postgres experience with a coworker,
I noted something weird in a query plan of my
 first post .
Editors 
 Of course as a programmer, I use some editors.
My most used day-to-day are VSCode and Zed .
VSCode extensions are really great, but when I need something really snappy, Zed is now the default.
It has also gained a lot of possibilities recently, with the Debugger and Agent tooling.
Key takeaways 
 JSON is more efficient than JSONB when only reading data 
 JSONB is useful for data processing, when filtering data on the database. But in this test, the higher cost of creating JSONB is not worth it, and there is also a price to convert it back to text.
In my previous article on aggregations on Postgres , we explored different ways of building SQL queries that join several tables.

 This article has a follow-up , where I fix a mistake in the subquery benchmarks and explore disk usage under production-like constraints.