Gold plating, in my mind, is all the programming which isn't strictly necessary, but is either 1. fun, 2. interesting, 3. extends capability for better insight, or 4. some combination of two or all of these. Gold plating can be expensive. It takes engineering time away from paying products, an immediate payoff, for hopefully longer term benefits in domain knowledge, engineering expertise, and other difficult to quantify benefits. However, as the cost of inference plummets, directing an agent to thoughtfully "gold plate" an implementation suddenly makes sense. As an example, I had an agent implement a manual evaluation prompt useful for verifying output from a DAG processer against the DAG defined visually in a PDF file. In the past I would have called it done with the specs. Now, for very low cost, I can run an evaluation through the DAG from the command line, feeding it strings and getting a path and result in return. Pure gold!

Technical debt as it's commonly referred to really is a misnomer, at least from a programmer's point of view. As programmers are hired for the most part to solve problems unique to a business, or within a business context, they are most often solving a problem new to themselves. This is why agile methods are so effect (a different, much longer conversation). Highly experienced engineers can draw on knowledge of related solutions. The point is that this gap - debt - exists in a large part because the implementation in that gap is novel. As a result, closing the gap gets punted into the future, inducing technical debt.

I've read - for decades - how software at least superficially resembles construction work. But that's not true. In construction, hundreds if not thousands of years of practical experience are leveraged, allowing a high degree of planning to be successful. In software, we often don't know exactly what we want to build, nor exactly how to build it. Which, in the last few weeks, has changed. As Andrew Zigler explains in Mise en place for agentic coding, we really can plan in advance, when our software agents leverage billions of man hours of experience.

I used to build and maintain CI/CD pipelines in various tools, most commonly spread between GitHub Actions, CircleCI, and Semaphore. These pipelines need to be maintained. APIs and a authentication methods evolve. However, it's time consuming to keep them all running, especially after projects have been dormant for a while. Thus I'm unhooking them. The AI/Agentic power curve is moving too fast right now, none of my personal projects are sensitive, and anyone I work for will have these tools already in place, or I will build them as part of the role. For now, there is no justification to continue maintaining them.

Having started with ChatGPT back in the hoary old days of 2023, I would think I'd be quite acclimated to the pace of change. But I am not. Change seems to be accelerating at an increasing rate. Now with the ability to spin up entire applications, I had a latent, perhaps subconscious impression that I would finally get a grip on my personal projects, get some focus, finish and ship. That is not the case. Instead, I've started many new projects, and not really shipped any of them. I may need to redefine what "shipped" means.