Articles, notes and updates on the agent, the tools, and the craft of building software with AI.
One cozy window for all your local projects.
Two releases ago, we added Claude Opus 4.8. One release ago, we fixed how it thinks. Today, we let you control how hard it thinks — all the way up to a new setting called ultracode.
Anthropic published @anthropic-ai/sdk@0.100.0 to npm at 4:49 PM UTC today.
Most AI coding agents give you a config file and wish you luck. Elyra gives you a searchable settings menu you can open mid-conversation without losing your place.
Here's a workflow I see constantly: developer gets a Slack message about a bug. Opens YouTrack in the browser. Reads the issue. Copies the description. Pastes it into the AI agent. Gets a fix. Goes back to YouTrack. Updates the state. Adds a comment. Closes the tab.
No new features. No new tools. No new commands. Twenty fixes for the things that were quietly wrong under the surface — the three-second freeze, the silent failures, the slow leaks, and the crash that breaks your terminal.
There's a moment every Laravel developer hits when working with an AI coding agent. You ask it to set up Sanctum token authentication, and it generates code for Laravel 10 when you're running Laravel 13. Or it invents a Livewire method that was renamed two versions ago. The agent is confident, the code looks right, and it's subtly wrong.
You're in the middle of a refactor. The agent is reshaping the auth module — renaming files, updating imports, chasing references. It's the kind of work that needs a few turns to finish. You're watching it tick along, and then a small thought lands sideways:
Most of Elyra is natural language. But four single-character prefixes — !, !!, /, and @ — give you precise control over what the agent sees, what it doesn't, and how you steer the tool itself.
Every coding agent can write Laravel code. You ask for a model, you get a model. You ask for a controller, you get a controller. The syntax is correct, the imports are right, and the code runs.