Articles, notes and updates on the agent, the tools, and the craft of building software with AI.
There's a pattern that emerges after a few weeks of using a coding agent daily. You catch yourself typing the same instructions over and over. Not the same tasks — the tasks are different. But the framing around them is identical.
There's a workflow that every developer does dozens of times a day. You're in the middle of building something. You need to check what's in the database. You open a database client — TablePlus, DBeaver, phpMyAdmin, the MySQL CLI. You write a query. You read the result. You go back to your editor. You continue coding.
Elyra ships with six themes — dark, light, palenight, catppuccin, onedark, and norwegian-midnight. That covers a fair range of moods, from the airy daylight of light to the deep blue dusk of norwegian-midnight. But "a fair range" is not "everyone's range," and sooner or later somebody asks the inevitable question: what if none of these are quite right for me?
The flashy features get the blog posts, but the changes that actually determine whether someone keeps using a tool happen in the first five minutes. Here's what's new in Elyra v0.6.1 and v0.6.2
Elyra v0.6.0 isn't about new commands or themes. It's about what happens inside the agent over the course of a two-hour session — how it remembers, what it costs, and which model does the work
Elyra's footer packs a lot of information into two lines. Token counts (input, output, cache read, cache write), estimated cost, context window usage with percentage, auto-compact status, model name, thinking level, provider name, git branch, working directory, session name. It's all there, all the time.
There's a quiet little failure mode that shows up the moment you bring an AI coding agent into a Dockerized project. You ask it to run the migrations. It cheerfully fires off php artisan migrate on your host machine — where there isn't any PHP, or worse, where there's a stray local MySQL that has nothing to do with your actual app. Nothing happens. Or something happens, which is somehow worse.
There's a particular dance that terminal-based development teaches you. You ask the agent to nudge a button's padding, or make a gradient less aggressive, or fix that weird thing happening to the heading at tablet width.
There's a particular kind of nervousness that comes with pushing a new package to npm. You hit publish, the tarball goes off into the world, and then… nothing visible happens. No fanfare. No counter ticking up in real time.
Or: why one tired agent is worse than five focused ones