to build is easy; to discern, divine

before november of last year, my usage of coding agents was minimal and quite utilitarian. my day to day work didn’t involve me actively writing code and that had been the case for about 3-ish years. there was the occasional need to roll up my sleeves to update niche serverless functions but for everything else, delegation was the order of the day. my experimentation with ai-assisted coding stayed confined to low-stakes personal projects. these were mostly small services deployed within my home network for various use cases, or the odd housekeeping script to automate what could otherwise be an inconvenient task.

my ide of choice for the last 10 years had been vs code. i got myself a cursor subscription in march of 2025 and installed claude code on the 2nd of the following month. agentic software development was shaping up to be a compelling concept and it helped to have a front row seat to observe where things were headed. i remember the first time it occurred to me that since claude code ran in the terminal, i could use it to reason through and perform operations directly on the machine it’s running on. i installed it on one of my windows desktops and used it to debug something that had bugged (hehe) me for some time and oh, what a neat party trick, i thought.

then came november. i usually take the week of my birthday (nov 12) off work so i was looking forward to spending the time stretching and crystallizing some of the theses i had formed over the past few months on how best to steer models into working for you. i planned to push the frontier models available at the time to their limits by having them autonomously build a small but well-defined product. my tool of choice was cursor and i was the hard-to-impress judge and jury.

as luck would have it, november became the month of interesting model releases. gpt 5.1 on the 12th, gemini 3.0 pro on the 18th, and opus 4.5 on the 24th. it didn’t take long for opus to get me hooked and start compulsively searching for things i could create with it. funny enough, it was the reception and perceived superiority of gemini 3.0 that caused sam altman to pen a “code red” memo at openai. quietly, opus 4.5 (along with its purpose built harness) will eventually go on to be a developer favourite over the holidays.

[obligatory llm-ism incoming]

claude code was no longer the stuff of party tricks. paired with opus 4.5, it was effectively a co-developer that would reason through problems you present to it and come up with working solutions within the bounds of your stated parameters.

it’s 2026 and everyone is back to work. they can’t stop talking about what they built with opus 4.5 over the holidays. even c-suite folks had gotten in on it, with many reporting that they had built functional apps for personal use in minutes. wait, c-suite? yes. everybody (product managers, designers, even marketers) was a developer now. a chasm had been bridged and simply being able to describe what you want to create was enough to produce a working output. writing code had been effectively democratized.

much has been said about what this means for the future of software development as a craft. from the hypeman’s favourite assertion that “saas is dead”, to some people declaring the redundancy of the human planning phase of software development altogether. one thing is for sure, the landscape is shifting in ways that we didn’t anticipate just four months ago. the model releases have kept coming too, with each lab pushing new versions in the months since. some are also building interfaces that bypass code entirely, putting value creation capabilities directly in the hands of non-developers.

i came across an argument last week which stated that the token cost to build a production feature is now lower than the meeting cost to discuss building that feature. its conclusion was that the planning layer is now a bottleneck that should be eliminated. “just build it in 2 hours”, it said. “measure it with a group of customers, and decide to kill or keep it.”

my initial reaction was to take the argument to its logical conclusion, where it eventually falls apart. your measurements will eventually generate false signals since the market will be inundated with a never-ending stream of half-baked products, thanks to everyone else having access to the same stochastic generators. the democratization of building is simultaneously a democratization of convention. push thoughtfulness aside, and customers get the median solution rather than one informed by a deep understanding of their operating context.

in the before times when building was expensive, planning served as a natural filter. not everything got built because resources were finite, and that constraint forced prioritization, which in turn forced thinking. in a world where planning is skipped, the forcing function for discernment is eliminated and so is the mechanism for ensuring you are moving in the right direction.

there’s also the matter of hidden costs. for example, we will see companies forget they are supposed to build for their customers, and will instead engage in side quests like rebuilding critical internal tools, lured by the promise of easing subscription fees paid to third parties. to build is easy. now you have a suite of unrelated tools to maintain and continuously improve. never mind the reality of non-obvious edge cases, architectural and security considerations, compliance, infrastructure, and the nightmare of retraining internal users. turns out the ingredient that makes software actually useful over time didn’t get democratized.

we are still in the early stage of this paradigm shift and it would be intellectually dishonest to pretend otherwise. the models will continue to get better and the tooling around them will mature. the workflows we are cobbling together today will look primitive in hindsight. amidst the noise, knowing what is worth building in the first place has always been the hardest part. the vast graveyard of companies that built relentlessly, but either failed to reach market fit or were unable to outpace their customers’ ever-present desire for something better, should serve as a lesson.

the willingness to sit with a problem long enough to understand it before reaching for a solution is your superpower. don’t let anyone tell you otherwise.