I'm Scaling Back My Agentic Coding and Going Back to the Basics

I’ve gone from being a total skeptic — believing these LLMs were fancy Google searches — to writing all of my code entirely with an agent, to now significantly dialing back my reliance on these tools. This is my why.

I remember when I first started using ChatGPT, maybe a little more than two years ago at this point, I thought “what’s the point of this?” To me it seemed like it was simply a better Google search. Before the tool I was doing what we all did: Google search a problem, get a StackOverflow answer, and copy/paste/adapt the solution. This worked great for nearly everything. That took care of your boilerplate code and weird loop/logic cases that you’ve never seen before. But then I started offloading that search work to ChatGPT and that did a stellar job of finding me the answer, and it wasn’t long until I realized that I can just copy/paste my code into ChatGPT, get it fixed/updated, and paste it back into my editor.

Now we have tools like Claude Code, where you can open the agent in your repo dir, give it a prompt to plan a feature with X, Y, Z constraints, have Claude investigate and generate a good-looking plan for you, and then run off and generate good code. It works great, it’s faster than me, and it just works. I’ve gone all-in on using Claude Code for 100% of code generation for the past 9 months and I’ve decided that it’s time to scale down my usage of these agents.

I feel worse at coding than I was 2 years ago, despite having 2 years more experience. I’ve offloaded so much of my thinking to a tool, that I’ve atrophied my coding muscle. The perfect analogy is that working your brain is just like working your muscles at the gym. If you stop exercising, you lose it — and I’m feeling it now.

With the amount of code you can generate, the number of features you can complete, the number of bugs you can solve, all in the amount of time it used to take to do a fraction of that work, it’s hard to not use the tool. Companies have championed for their IC’s to massively increase their usage of coding agents so that they can work faster. The idea was great, but I think the atrophy of skill, agent token costs, and what most engineers actually do while the agent works (the human scrolls on their phone — they just traded their thinking for token burn) has been a net-negative.

Additionally, why am I relying on my subscription to a tool to do my job? I can’t run a trillion-parameter LLM on my laptop. I never will be able to, so obviously you have to pay for a subscription to use it. Okay great, that’s fine, we all pay for subscriptions to tools, but these ones are different. When you rely on it to entirely perform what was/is a massive part of your occupation, it feels wrong. The idea of not being able to work offline (say from a plane mid-flight) feels horrible. My ability to work should not be dependent on Anthropic’s service availability or my connection to the internet. I remember several times where Claude was down for a few hours — I felt like I needed to be paying for multiple agent subscriptions for high-availability. There are no other tools in my belt that require HA to do my job. That’s a personal vulnerability in the sense that I am massively less effective without my agent to think for me.

The argument in favor of heavily relying on agents for coding has been that your customers aren’t paying for handcrafted code — they’re paying for a result. So why should you care how that result is generated, whether it be by an agent or a human? I say that you, as the engineer, should care because codebases that are mostly or entirely generated by agents create a comprehension gap. As more and more code is not generated by you, or written in your style (which is really a projection of how you think), you become less able to understand it. The agent does, and the agent writes it in a way that it understands, but perhaps not how you do. Some folks are saying code reviews are the answer. I say they’re not. Code reviews have always sucked. The git diff has always been a terrible way to comprehend the changes being made. Sure, they show you what changed, but its hard to really give meaningful input regarding the broader effects of those changes when you just have a diff. Especially at the volume that agents produce code at, CI checks become the real test of whether to merge or not merge, but that doesn’t catch everything, and neither the git diff or the CI checks improve your repository comprehension.

I’ve tried the hyped-up agent loops from a task queue which promised overnight “code factories”, crafting PRDs with the agent to make sure it builds exactly what its supposed to, designing state machine outlines for the agent to follow so that it can built the PRDs without straying from the system design. I really feel like I’ve tried all of the things you should to get results. Like I said, the agent writes great code, but its the things you find when you actually dig into it while simultaneously understand less and less of the code as it’s built. You can have a code factory, but now it’s not your code. If you go to debug it and understand it without help of your agent, I think you’d find it an increasingly difficult task. When your production app runs into a sev0 bug, you better hope your agent provider isn’t having an outage.

When you write code by hand, you build systems. You make sure your code is DRY. You can tell if two functions are so similar they could probably be collapsed and simplified. You are actively thinking of “how can this break” or “am I doing things securely”. Unless explicitly directed, I’ve found my agentic coding to write 3000+ lines of code without breaking up files, because I didn’t watch how the file was growing over time, as I wasn’t directly in the code base. I’d have the agent build design or logical systems to reuse across components and pages, but it would sometimes ignore those and invent one-offs for things it was doing. We’d solve a bug, and then introduce 3 more because it solved for that bug without critically thinking about how that affected the other systems. You could say that’s an error in my usage or dilligence to review code, but I’m not special, so I’m know this is happening for everyone else whether they’ve realized it yet or not.

With all of that I’ve said here, I’m scaling down my usage of agents for coding. It feels fast to generate hundreds of lines with prompting, for a while it felt awesome to be able to write code without staring at a screen all day, it feels productive to go from 0 lines of code to a mostly-functional application in a day. The tax comes after that initial project development at the cost of skill atrophy, loss of functional independence, depletion of repository context, and the constant iteration loops you end up in trying to wrangle the agent to fix an issue right or build a new feature correctly the first time (in one session, not a one-shot prompt). I’ve been beaten up paying the agent tax, and now I’m going back to basics. Agents can write boilerplate, help me think through things, and help me set up projects, but they’ll no longer be driving development. They are simultaneously a thoughtful peer, and a junior engineer that needs tons of supervision and coaching. I think it’ll be more productive to go back to doing things myself.