I agree. It really is not a difficult “skill” to learn. It took me probably 4 days to configure agents to break down requirements, write tests, write code, open PRs, review and merge PRs. Learned how to use skills, MCP, AGENTS.md. It is really not complicated. It’s just…writing well. Knowing how to decompose problems. Trying out different tools. I also learn new things every day but I sincerely do not think I would be X times more effective with these tools had I started a year ago.
Mix in exhaustion from previous hype cycles (NFTs?) and folks pushing the narrative that if you currently write any code manually, you’re obsolete and it’s hard to not want to push back negatively. Even if you find the tech interesting/useful.
It’s a tricky philosophy to put into practice. I have oscillated between this approach (“owning” the effort) and “owning” the outcome. I have found that taking ownership of the outcome leads to better results because I have a personal stake in the outcome and I tend to think through the problem more deeply, but I am almost always left feeling more stressed and “empty” when the work is finished. When I focus on doing the best I can and let go of the outcome, the end result is almost always subpar which leaves me feeling frustrated, because I know it could have been better had I taken on more responsibility.
Coaxed into dead-end architecture is the exact issue I have had when trying agentic coding. I find that I have the greatest success when I plan everything out and document the implementation plan as precisely as possible before handing it off to the agent. At which point, the hard part is already done. Generating the code was not really the bottleneck.
Using LLMs to generate documentation for the code that I write, explaining data sheets to me, and writing boilerplate code does save me a lot of time, though.
I appreciate your reply. My tone was a little dismissive; I'm currently deep deep in the trenches trying to unwind a tremendous amount of LLM slop in my team's codebase so I'm a little sensitive.
I use Claude every day. It is definitely impressive, but in my experience only marginally more impressive than ChatGPT was a few years ago. It hallucinates less and compiles more reliably, but still produces really poor designs. It really is an overconfident junior developer.
The real risk, and what I am seeing daily, is colleagues falling for the "if you aren't using Cursor you're going to be left behind" FUD. So they learn Cursor, discover that it's an easy way to close tickets without using your brain, and end up polluting the codebase with very questionable designs.
It's 60-80% as good as Stack Overflow copy-pasting programmers, sure, but those programmers were already providing questionable value.
It's nowhere near as good as someone actually building and maintaining systems. It's barely able to vomit out an MVP and it's almost never capable of making a meaningful change to that MVP.
If your experiences have been different that's fine, but in my day job I am spending more and more time just fixing crappy LLM code produced and merged by STAFF engineers. I really don't see that changing any time soon.