AI Coding Thoughts 2026

13 Aug 2026

Recently a video popped up in my feed about a developer’s experience of coding with AI and why they were done with it. I found it interesting and began writing a comment to respond to it, only to realise it became rather long. So I thought I’d expand it into a blog post about my thoughts on coding with AI, LLM-based agents etc.

First of all, the video in question was one from the YouTube channel Brett Codes titled “I’m done coding with AI”:

I encourage you to watch it first to put my thoughts in context. I’ve also used a tool to summarise the video, but these summaries are never as good as watching the whole thing. Brett also wrote a blog post related to the video, but I am addressing the video content rather than that post.

First of all, his early experience with the Copilot’s autocomplete function partially mirrors my own experience with some AI-based tools. They can be distracting and not useful when you’re in a flow state. We had mostly learnt this lesson with auto correction software for text, but in recent years it seems to have been unlearned.

Previously, it was normal to see a red or green underline in your editor. You’d later correct the mistake, or go back quickly and correct it as you wrote. We then switched to autocorrect fixing common typos, It works well most of the time, when it is correcting simple things, but can be annoying when it overreaches. Now we have editors suggesting multiple words and phrases, robbing you of the thought process that goes into writing and/or biasing you towards more sterile language. I know in recent emails I have sent that I got very annoyed at the suggestion of how I should sign off. I will choose when I am being sincere, thankful or curt.

Brett also describes an incident when debugging an issue in 2022/2023 related to an AI hallucinated software dependency version. I’ve experienced something similar even going into 2026. Large language model training currently cannot stop such hallucinations, and this is a fundamental issue with the architecture they’re built on. Much to the chagrin of AI evangelists, who’ve claimed that hallucinations no longer occur.

I am sure hallucinations happen less frequently than before, thanks to better tooling providing more live context, but that doesn’t prevent them. This is a main reason why you should always be sceptical of the output of AI. It can confidently “cite” sources which, when you come to read them, don’t actually say what the AI produced text claims they say. The issue the evangelists have is something akin to Gell-Mann Amnesia: they believe the bluster from the AI because of the source and their own limited knowledge of the subject. So when AI makes a mistake they either don’t realise it or discount it as something minor.

With all that said, it doesn’t mean AI isn’t a useful tool. For example, summarising documents you’ve written or read is a task that a large language model is well suited to do for you, since you know the contents. I know I’ve had limited successes with them summarising meeting transcripts, provided I either prompt it with my own notes or steer it to the items of importance.

Brett mentions a health incident where he erroneously went to the emergency department based on the advice of a chatbot. While I hope he is well, this isn’t that dissimilar from people heading to see doctors after Googling their symptoms or looking them up on WebMD. It was definitely a good wake-up call on how he should doubt AI-generated output and always remember that he isn’t actually communicating with a reasoning being when engaging with a chatbot. Again, AI isn’t completely useless here. Provided you give it enough context and take what it says with a grain of salt, it can help narrow down a potential ailment based on symptoms so you are better prepared when seeing a doctor. But you have to always remember that there’s no substitute for a medical doctor’s training and experience.

AI chatbots, more often than not, act more like a mirror when you present them with emotional language or leading arguments/phrases. This isn’t that surprising when you understand how the raw model works after pre-training. Next-token prediction will naturally try to continue on the given context; that’s fundamentally what it’s designed to do. Even with post-training, this behaviour still exists; it’s just mixed in with the question-and-answer format.

He also mentions the pressure to use AI. This is something I think most people in a lot of industries can sympathise with, and I think he does identify one of the key reasons why: pressure from higher-ups who’ve bought into the hype from AI providers.

In the second half of his video, he describes his experiences using AI coding tools and agents. The existential dread, building becoming review, and distance from the creative process reducing care, are all things I can sympathise with. Again, I recommend you watch the video as I don’t want to retread all his points in detail.

The feelings of dread and purposelessness are highly relatable, and I think stem a lot from the hype and discussions around AI replacing creative work like software engineering. We’re however many years into the “AI will replace coders in 6 months” era, and that constant uncertainty is sure to play on anyone’s mind, even if they don’t fully believe it. We’ve seen claims around 100x improvements in output, people proclaiming themselves to be productivity wizards with 5+ Anthropic subscriptions, “autonomous” agents hacking websites, and dozens of security vulnerabilities revealed by AI models. It’s enough to make you want to cut your internet lines and live in the woods.

The things we need to remember about all this are threefold:

1. There’s a lot of marketing hype from these AI companies, chip manufacturers and infrastructure providers. They have clear monetary incentives to get you on board with their hype train, either in support or in opposition. Either way you’re thinking about them and/or spending money on them.
2. We’ve not seen these supposed gains from AI. At least not consistently, especially not the outrageous claims that are often thrown about. There’s been no obvious economy-wide explosion in productivity or in the number of successful businesses built primarily around AI. Sure, it’s being used successfully to some extent, and there are businesses being built around it, but the jury is still largely out on these businesses since many are still in their early funding stages.
3. Any investment on the same scale as what AI has received in the last few years would inevitably have produced results like those we are seeing. Take the security vulnerabilities found: had the same amount of money and effort been spent on security research, it’s likely that we’d have found a similar number of vulnerabilities.

I’m sure I could make other points too, but the above three are hopefully enough to fight off the dread and ennui.

Brett’s other concerns around the work becoming tedious, the impact on caring, learning and quality all stem from what I’d argue we could call misuse of AI. Not Brett’s misuse in particular, but the wider industry and attitudes towards its use, and the patterns of use being encouraged.

AI powered by large language model is still a tool at the end of the day. I am often reminded that “a bad workman blames his tools,” normally when I am yelling at my computer about a mistake that is inevitably revealed to be my own!

We should be using AI to automate the parts we don’t like out of the process, not letting it take over the things we enjoy. I know an early experiment with AI-generated code in my RPG Party Members project yielded workable code in one of the interfaces, but it ended up being code I did not wholly understand. It robbed me of the pleasure of learning how to do something myself, while implementing something close enough to what I was imagining.

I can and do use LLM tools like Codex to generate unit tests for code I have written. The generated tests run and essentially work to ensure the current functionality doesn’t change over time. This is similar to how you write tests in non-test-driven development. I can still write said tests myself, and sometimes do, at least initially, but they’re tedious and not the area I want to spend time and focus on.

Of course, with my generated tests I have to understand that the generated code is neither infallible nor revolutionary. More importantly, I need to review these tests to make sure they’re valid.

Has the AI removed some benefits and increased my distance from the code? Probably. The issue with getting a coding agent to work for you is that it removes you from the coding feedback loop. Every moment you are coding, you are making micro-level decisions, you’re solving problems, and evolving your understanding of the wider software system being built. Even when writing tests this is true; I might spot a logic error, typo or other nonfunctional issue for example. These are things that AI-generated code just breezes through or makes assumptions on your behalf about.

This micro-decision problem is magnified when you ask a coding agent to do non-simple tasks. Suddenly, you’re no longer involved beyond the occasional prompt. Your understanding of the software doesn’t become much deeper than the high-level design, and if you want deeper understanding, you then fall into the AI code review trap. Either you review every statement and internalise it, or you rely on the AI to have been correct.

When reviewing code written by a fellow human you can rely on them since they have an ego, a level of experience, competence, their own style, and an internal model of the software in their mind. You know that Bob writes code a certain way, has made certain mistakes in their past, and is less likely to make them again, so you can breeze through certain parts of a code review as a result. A coding agent has no ego, making a mistake to it doesn’t cause it anxiety, it doesn’t bear the risks, it can simply vomit out code that is statistically “good enough”.

So what I am saying is, we have misaligned our use of these AI tools and that is causing a lot of the friction and woes Brett talks about in his video. If any higher-ups in business are reading this, I implore you to not mandate specifically how your teams use AI. Instead, you should empower them to use it within limits, remind them to use it to smooth over the parts of their work that are more mechanical and frustrating. Don’t tie metrics to its direct use, that’ll create perverse incentives and more than likely run up your bill while making some people miserable, and risk deskilling your teams, forcing reliance on the AI providers.