It’s your own context window that isn’t enough…
LLM context windows have grown from thousands to millions of tokens in the span of a few years. The best coding model I’ve used — Claude Opus 4.5 — accepts “only” 200k, but it works so well, that if you keep your project metadata tight, that’s enough to hold architecture, conventions, and plenty of code for even large codebases.
IMHO, when you know “how” you want something to be done and you specify this in a great plan — it does so well, that the limitation is no longer the tool — it’s the one who wields it.
Up until recently, I’ve been teaching developers to focus on 1 task for longer stretches (4–8h) -> get into flow mode -> jump over walls that would otherwise take tens of 2h blocks (or even be impossible to do) if you switch context often. But now the game has changed…
We are fundamentally changing the way we build software. I’m sure a bunch of people will immediately jump at this line, but I’m typing code less than 10% of the time I’m coding (could be 5%…). Just being honest here. Implementation details are now so well covered from the coding model that I can point to our conventions, project meta, I pass in architecture in the planning stage -> all fits in the context window + some gymnastics on compacting it during higher count of iterations -> it figures out the trivialities and follows my instructions well.
But that’s not enough, because now you get those gaps where your agent is working and you are waiting for results — here comes parallelisation. Parallel agents in tools like Cursor still don’t work great (which will change super-fast), but they already allow you to split into doing multiple tasks at once. And here comes the problem for all of us with 20+y of experience in coding — we are not great multitaskers. In fact we have intentionally built a mindset of working on a single task at a time.
I believe this got us here, but it will not get us to the next level. You are no longer the actor responsible to figure out the lowest level of detail and when working on higher level of abstraction — you can (and should) successfully multitask. Everyone who manages multiple entities producing intellectual output — can’t work sequentially. You become more of a “very involved architect”, and less of a “coder”.
So right now, I battle a cognitive dissonance I experience while multitasking in Software Engineering and I am starting to feel a huge anxiety when stepping away from my station, being unable to unblock my agents to continue working.
What does your process look like? How do you make the best use of the incredible tools we now have at our disposal?
