My position is that “relative ease” for anything but common or non-complex functionality currently means (for example) I had to spend 2 hours yesterday explicitly and repeatedly telling the agentic coder to exactly replicate the medical UI for the v6 file version or our practice monitoring platform despite my having previously implemented and perfected the v5 version over the last year so it had existing examples to follow. The two hourse were spent to double-checking and repeatedly pointing out that it had implemented absolutely nothing the same as the previous version (including randomly changing the patient selection criteria !?!) - for which it apologised profusely but only made the modifications with explicit identification and repeated explanation of what it had still done wrong every loop (and it would totally drop the ball if I gave it more than one at once). And this despite me having explicit specifications, deterministic evaluation scripts, and screen based checking. Agentic coders remain unreliable for implementation of complex functionality unless you have a human double-checking absolutely everything. Thus if you are doing anything which requires complexity, reliability, or highly specified operation (eg most clinical tools), they still have a long way to go. This is probably marginally better than 12 months ago, but is definitely not the orders-of-magnitude improvement I had expected over the past 12 months. And this is essentially the same whether I’m coding C++, Python, or a React Native app. For every 2 hours of agentic coding I have found it needs at least as much time running deterministic tests and then another hour of testing to ensure it does what the task/contract it was given explicity stated. And if you don’t watch it much of the time it’s coding it has a high chance of going off piste…
So to rewrite that component from scratch would likely have taken me 10 hours to handroll - but it still isn’t a negligible investment of time and expertise.
Similarly with a running coach app I’ve written for my phone - theres about a 30% chance that any new feature will result in it breaking some existing feature - and even with emulation and automation on a real phone running literally hours of simulated runs, problems still slip through and only become apparent during the live human-tested run.
