There was a two year period around 2011-2013 where I experimented with my dev team. We were being "forced" to migrate a legacy enterprise system from .Net/MSSQL to Java/PostgreSQL with replace the front end with a modern, reactive web interface. Only two of the existing developers had Java experience, and both were conveniently senior engineers in two different offices and running their own discrete sub-teams.

One of the guys had a very strong opinion that the ideal architecture was something as abstracted and object oriented as possible with single function classes, etc. I let him run with that. The other guy got frustrated with his sub-team's inability to write code to spec in a language they'd never used before and where they were trying to build some new features they didn't clearly understand. He developed a strong feeling that TDD was the most efficient path forward: he owns the PRD and design, so he just created test stubs and told the remote team to "just write code that passes the test" even if they didn't understand the function of the block.

So, after a few months where did we end up:

1. The "abstract everything" architect's team had an extremely fragile and impossible to maintain codebase because it was impossible for any outsider to tell what was going on.

2. The "just pass the damn tests" guy had a team that had quickly ramped on a new language and they had a codebase that was incomplete (because they were building it like a Lego project) but that everyone could understand because the code blocks generally stood on their own.

What was the next step: to shut down the guy who abstracted everything and force him to drive a quick & dirty rewrite that would be more maintainable, and to also start a major refactoring of the "Lego" team's code because it was so fragmented that it also was fragile and unsuited for production.

I saw this as a terrific learning experience for all involved and I was able to get away with it because the stakes were pretty low and we had time to experiment (part of the ultimate objective was upskilling the team), but the more important lessons were these:

1. Docs matter. Take the time to write clear & detailed specs first because you'll be forced to think of edge cases and functionality that you didn't originally, and it provides a basis for system design, too.

2. Architecture & design matter. Adhering too close to any single paradigm is probably a mistake, but it takes experience on the team to understand where the compromises are and make the best decision for that system.

That second point will not stop being true with the advent of agentic assisted software development. Like others have said, my expectation in the job market is that pay will continue to be depressed for junior hires as employers reset expectations and generally just want folks who can instruct fleets of agents to do the actual coding. Senior staff will become increasingly critical and their jobs will be painful and difficult, because it'll be assumed they can (and will be willing to) do design & code reviews of artifacts originated by agents.

What I am going to be most interested in is what happens in the SRE/Sysadmin world over the next few years as more AI-generated code hits prod in organizations that don't have adequate review & oversight functions.