Agreed!

I had a calendar reminder to write a sequel to this for 2026 and still intend to do so.

Here's a preview/draft/spoiler:

The way I'm using AI now is different from what it was a year ago. Yesterday afternoon, I used Claude Code with https://github.com/obra/superpowers, https://github.com/intersystems-community/iris-dev MCP, and an internal skills library to do 90% of the implementation converting an Excel spreadsheet to IRIS persistent classes, REST APIs (via https://github.com/intersystems/isc-rest), and a few Angular components, with data load and change control integration. I iterated on and reviewed the design and implementation plan, then it chugged along in the background while I did other things. I kicked the tires, shared some screenshots with a pleasantly-shocked stakeholder, did a few hands-on-keyboard fixes for the data load after reviewing the code, and some additional iteration with Claude. (It's better at CSS than me. I hate CSS.)

This morning, I've done a few more things hands-on-keyboard that are quicker to do myself than to explain to the LLM (though we do have a skill that helps, there are still has some time- and token- intensive steps). As I eat my breakfast, write this comment, and respond to some emails, Claude is working on the last 8%. If I'd done all the work myself, this project would have taken 3-5 working days of "wall time" with fragmented attention between meetings and other tasks - which, practically speaking, means it wouldn't have happened; I would have built a simpler interim solution to meet a deadline and my stakeholders would still be satisfied. Now it's down to <1 working day, and I get to delight my stakeholders instead of just satisfying them.

My expertise in IRIS is still needed to steer the agent correctly. I spend more time on architecture and design than easy mechanical coding. My hands-on-keyboard development time sometimes goes into tricky problems the agent can't solve on its own, or - when it gets things wrong - into building skills and other feedback/feedforward mechanisms to help it not make the same mistakes again. And most importantly, my job is still fun and fulfilling - my creativity goes as input into the agentic process and, at a higher level, into making the agentic process more repeatable.

As a software developer, I don't delight in algorithms, academics, or rote coding. I delight in building things that help people. AI makes that faster and easier.