非AI公司CTO預測2026年軟體開發趨勢

Behan’s Substack

Reflections on the best year of my life and bold predictions for 2026

In 2025 sci-fi level technology arrived. In 2026 it'll go mainstream.

Year in Review 2025

2025 was the best year of my life. I got married to my best friend, bought a house, and spent 3 weeks traveling across China. It was also the most productive year of my professional life. Thanks largely in part to my second best friend (Claude), I shipped significantly more features, wrote more documentation, and implemented more test coverage for Moovs in the last 6 months of 2025 than the previous year combined.

After my experiences this year (and the 27 years prior) I have a pretty strong conviction that I am in fact the luckiest man alive, and possibly the luckiest man to ever live. I get to spend my days solving problems and building with people I care about. I have an abundance of meaningful work; I build software that empowers over 1000 transportation businesses to be more effective. And an even greater abundance of meaningful relationships, both personally and professionally. I like to imagine that if Jeff Bezos could peer into my consciousness, he’d desperately try and trade lives with me, offering up his super yacht and all. But to give myself credit, despite being the luckiest man alive, my current position in life isn’t entirely luck. I work hard, have difficult conversations, show up everyday, dare to try, fail often, and reflect.

“The future is already here, it’s just not evenly distributed” - William Gibson

I had a moment during a work trip to Palm Desert, California a month ago that I feel perfectly sums up the state of 2025. Several of my colleagues and I who live throughout Canada and the US were in a Cybertruck that was self-driving us to Joshua Tree whilst we were learning Mandarin from Grok in voice mode and generating photos of ourselves with prominent celebrities and political figures using Nano Banana Pro. If I were to show a video of the scene to myself from just a few years ago I would have dismissed the whole thing as science fiction.

We live in a world with self-driving cars that can take you driveway to driveway, change lanes, stop for pedestrians, merge onto highways, and brake to avoid reckless human drivers. We’ve developed LLMs that can ace the Turing test and no one seems to care. We constantly shift the goal posts about what constitutes “AGI”, meanwhile modern frontier models can act as personalized language tutors in just about every major spoken language but this isn’t “general intelligence”? State of the art text to speech models sound indistinguishable from real people, intonations and all. Image models are so good at generating and modifying images that you can no longer distinguish fact from fiction. In 2025, sci-fi level tech arrived. My prediction is that in 2026 this tech will go mainstream.

How my software development workflow has changed

In 2024, my software development workflow looked like this: Half my monitor was dedicated to Claude, the other half was Cursor running Claude 3.5 Sonnet. I would plan, ask questions, and ideate with Claude, then implement features by hand using Cursor + the tab auto-complete. For extremely well defined tasks I would have cursor’s agent mode take the first crack, after which I would generally have to refine the output by hand. Unit tests were another area that I could have an LLM write mostly by hand. This resulted in about a ~50% speed up from the pre-LLM era of building software.

In 2025, my software development workflow looks like this: I have 3-5 desktops each with 2-6 terminals open running Claude Code with Opus 4.5. Moovs has separate repos for the server, operator web client, customer web client, and driver mobile app which means I usually have a separate terminal running Claude Code from a git worktree for each repo, although I’m experimenting running a single instance of Claude Code from a parent folder that has each repo within it. I start each task by writing a concise plan in Claude Code’s plan mode, then iterate a few times on that plan before I set Claude off on auto-accept mode to implement the feature in its entirety. For smaller features or changes, Claude usually gets it right on the first try. For medium to large features I first focus on getting the feature working, not caring at all about the code produced. I then spend time improving the code and reducing the size of the change as Claude tends to be overly verbose. I repeat this process for 2-6 features at once, where each feature usually touches at least 2-3 of Moovs’ repos, meaning 4-12 instances of Claude running simultaneously. Before putting up a pull request I review all changes locally with both Claude Code and Codex and address all real issues before requesting a review from my human teammates. This has resulted in a ~4x speed up from my development workflow of 2024. However, coder beware, I do feel like I’m slowly developing schizophrenia from all the context switching and multi-tasking that I’m doing.

The way in which I wrote most of my code 2 years ago (by hand) has become an antiquated practice reserved for intellectual stimulation, learning, or the ever-shrinking times where Claude can’t figure something out. It’s become so rare that my colleagues and I now refer to it as “raw dogging it”. If you’re a web developer (the domain in which LLMs have the most training data) and your primary goal is shipping product as opposed to programming for the sake of programming, having an LLM write most of your code is the likely future of how software will be built if you want to stay competitive.

Ideas for the future of building software

There seems to be 2 broad visions of the future of software development right now, one in which AI becomes ubiquitous for building software and greatly improves both the quantity and quality of that software. And one in which using AI to build software is a mistake, we build systems we can’t understand, tech debt accumulates beyond the ability of an LLM to manage, and the world’s software systems come crashing down, forcing a reversion to the good ol days. I find myself presently in the first camp but I empathize with those in the second.

In the search for which future world is more likely, it’s valuable to look at the attributes of the typical “AI-pilled” and “Anti-AI” developer and speculate “Why do they believe what they believe?”

From my observations and anecdotal life experience, AI-Pilled folks are by and large younger and less experienced than Anti-AI folks. Their age and lack of experience makes them more open-minded and willing to question the status-quo, hence their willingness to adopt AI. But their lack of experience and battle-scars also makes them more susceptible to oversights and false confidence. They have less experience with the treachery of complex code and may be putting too much faith in the ability of LLMs to manage that complexity. There may come a time when the complexity produced by their LLMs is too much for the LLM itself to manage, and because they are overly reliant on having the LLM write code for them, they can’t understand the mess they’ve built, and progress comes to a crashing halt.

Many strong proponents of AI-driven software development are also talking their own book. They work for or actively invest in AI companies and have a vested interest in the success of AI. There are also lots of hobbyist AI-Pilled programmers who work on toy projects with minimal to no users, which is a fundamentally different thing than building and maintaining large complex systems with real paying users.

However there are also counter-examples to these archetypes. There are strong proponents of AI with no vested interest in the technology’s commercial success. There are developers building real, complex systems with large numbers of real, paying users. I’m one such counter-example.

Maybe the experience of Anti-AI folks allows them to see risks and insurmountable obstacles that the AI-pilled folks cannot. Or maybe, the Anti-AI folks are just close-minded and intellectually conservative. Maybe their world-view of software and “The way things are done” has become sclerotic through years of repetition and stasis. Maybe they also feel threatened that their ability to write code, honed through decades of experience, is no longer as valuable as it once was. Maybe they identify too much as “Programmer” and not enough as “Problem solver”. Maybe they haven’t actually tried coding agents running the frontier LLMs so they don’t actually know how good the models have gotten and are basing their opinions off of the experience they had coding with LLMs a year ago. Maybe this change and threat to their identity and the value they provide is too much for them to bear, so they dismiss the entire technology as a fad, and continue to write code by hand with their heads in the sand.

My probabilities of each future coming true are about 80-20 in favor of the AI-pilled future. Here are my predictions for what both possible futures, AI-pilled and Anti-AI, will look like.

AI-Pilled Predictions

Most devs will write most of their code with some sort of coding agent.

Human code review will be completely replaced by AI code reviews.

The bar for shipping code will be so high that you won’t have time to carefully read all of the code you write. Instead we will normalize the practice of having AI write code for you, then have AI explain the code and visualize the added data flows, reviewing these descriptions instead of the code itself.

The work done by software developer will look more similar to playing Factorio than it will programming. The primary focus will be on building and maintaining scaffolding that allows coding agents to work effectively, instead of being the one to write code.

The job of software developer, product designer, and product manager will merge into a single role that has the responsibilities previously held by all 3.

Software developers and product designers, due to having more hard skills, will be better positioned for the convergence of the 3 roles than product managers.

The best developers, product designers, and product managers of today will still be the best builders of tomorrow. AI is a force multiplier, not a personality transformer.

AI-augmented self-learning will be widely accepted as more effective than traditional forms of education.

Students will begin choosing self-educating with LLMs and peer study groups instead of going to University, and those students will be more knowledgeable and hireable than their University Student counterparts.

Anti-AI Predictions

There will be substantially more outages from big software companies like AWS and Cloudflare due to increased LLM usage, AI Slop, and developers not understanding the systems they build.

The average developer will not know how to write a for-loop by hand.

The scaling laws, which say you can just throw more data and compute at LLMs to improve them, will pause and the capabilities of LLMs will plateau.

After having time to adjust to the now-capped capabilities of LLMs, we will realize they weren’t as capable as we thought they were.

Tech debt will accumulate across big tech-companies and AI startups to an inflection point where the systems become so complex that new functionality can no longer be added. Reliability and innovation at these companies will collapse.

Programming with zero AI assistance will come back into fashion. Companies will begin banning the use of AI and job postings will explicitly seek those who program completely by hand.

There will be a significant divergence between students and new grads who use and don’t use AI, with the those who don’t use it being far more capable of critical thinking and problem solving.

LLM usage will drop significantly and the AI Bubble will burst. Nvidia’s share price drops by 65%. OpenAI goes bankrupt.

And that’s a wrap. I wish all of my AI-Pilled and Anti-AI friends a Happy New Year and a prosperous 2026!

Thanks for reading my Substack! I hope to write much more in 2026. Subscribe if you’re interested in my takes for what is certainly going to be a bonkers year.

No posts

Ready for more?