Coding agents make it possible to do lots of work in parallel. I always have several (often many tens of) agents working at once, especially when working in Saggar. They might be working across many repositories, or in the same codebase. They write code, run the tests, and commit... then BAM!
I don't think the terminal is going to die.
I think it's finally going to become usable.
I've been building Saggar, a standalone terminal app shaped around projects rather than a pile of disconnected windows.
Nobody writes machine code anymore. For a good reason - it's inefficient to write. Luckily, we don't need to, because somebody already wrote it for us.
A lot of new AI coding tools are built around the same idea: start a thread, keep feeding it work, and let it run forever. You can see the assumption hardening into product features: Claude Code lets you name sessions, find and resume them by name, branch them, and export their transcripts. Cursor has editable chat titles, stored history, Markdown exports, and a way to pull old chats into the current one.
Spec-driven development means treating a written specification as the source of truth: define the intended behavior first, then implement and validate the code against it.
I find the principle intuitive and exciting, but it has the same problem we've always had with documentation: drift.
used to be fairly extreme about separation of concerns in my React code. A smart component file for logic. A dumb component file for rendering. A styled-components file for styles. An index file to tie them together. Four files per component, minimum.
It made each individual file very readable. You could generally fit one on a screen. If you needed to understand the render logic, you opened one file. If you needed the business logic, you opened another. The mental overhead of jumping between co-located files was low.
Claude Code's skills system is genuinely impressive. You can package domain expertise, best practices, and specialised workflows into reusable modules that extend what Claude can do. The architecture is elegant. The potential is huge.
There's just one problem: Claude barely uses them.
When an AI assistant explores your codebase, it reads files. Each file costs tokens. A lot of that reading turns out to be unnecessary—the AI loads a file just to discover it's irrelevant.
A small thing I'm pleased by: two plugins I made are now listed in the official Docusaurus docs.