March 21, 2026 · 2 min read
It's crazy how much work Software Engineers actually have to do such as frontend, backend, infrastructure, security, observability, scalability etc. That's already a lot of context switching.
December 2025 was when I was fully convinced AI could takeover a huge chunk of software dev. Ironically, a lot of dev skills can now be outsourced to coding agent skills. Forget the 10x engineer, the expectation is now a 1000x engineer.
Previously the limitation was that me the human could only work sequentially. Given 3 tickets, I'd have to pick 1 and finish that before moving on to the 2nd ticket, then the 3rd. This all happens sequentially. The context switching happens, when I'd have meetings, receive a Slack message of a different bug that occurred or have to reply to a client email plus the occasional Twitter doomscroll whilst waiting for a deploy to finish. Not so bad eh?
Now I have 3-4 agents running in parallel and in that time period, I am switching tabs every few minutes to check on the agent. When I change the tab from lets say agent 1 to agent 2, I have to quickly recall what I was trying to do with agent 2. By the time I'm done reviewing agent 4's code and back to check up on agent 2, me myself, my own context window hit capacity and inevitably compacted horribly to the point I already kinda forgot what I was doing with agent 2. And let's be honest, we went from checking Twitter while waiting for a deploy to checking Twitter between agent tabs too. Context switching used to happen in 10-30 min intervals but is now happening in minutes if not seconds. Yes, output has skyrocketed, but at the expense of extreme cognitive load and a shrinking attention span. And when agents spit out output at the speed of light, dopamine kicks in, and I'd be like why not run 1 more prompt. It's the same dopamine loop as online competitive videogames, just 1 more game! Rinse and repeat. I'm actually no different than a hamster on a wheel except the wheel keeps spinning faster.
You, the maestro of agents, gotta figure out your context window because unlike LLMs, your capacity can only be improved so much. It's not like you can subscribe to a 500 dollar super pro subscription to support a 1 million token context window. Because just like LLMs, once your context window goes through compaction, you're gonna make mistakes, and you can't afford to when you're shipping to production. Context switching overload was already a problem before the introduction of AI agents, but shit just got 1000x worse. To prevent burnout, context engineer thyself to know your limitations before you context engineer yo agents.