Let's start here,
Last Sunday, around 2 am-ish, I was talking to a friend, another engineer, about how AI has completely changed how we ship code in 2026.
I told him I had pushed out about 3 weeks' worth of code in the last 24 hours, and still had 6-8 different PRs and a couple of implementation plans from Cursor that I needed to review, steer, and approve.
Turns out, he was in a similar corner.
Curse of being pre-AI,
As engineers who wrote code pre-AI, we still kinda find it difficult to hand full control to an agent.
I still manually approve almost every command and review every line of code generated by an agent, and I have never allowed agents anywhere near my servers, even in staging.
In the last few months, coding agents have gotten so good that their output is almost always 60-70% what you want, and in some rare cases, you get 100%, but the non-deterministic nature of AI agents means that when they mess up, they do it phenomenally well.
I've interacted with other engineers who aren't pre-AI, or who started coding just a year or so before proper coding agents became good. (I'm not talking about vibe-coders who don't understand for loops) They are way more trusting of AI agents, and they ship much faster.
My trust issues have led to me being a blocker to the multiple threads of agents I have running simultaneously.
Ok, I'm the problem.
Today, I'm confidently 100x faster at shipping code than in 2019, and I recently built a project that should have taken a month or two, in less than a week.
I could probably be even faster, but I'd rather prevent the 10% chance of an agent dropping my entire database.
In 2026, the problem is no longer writing code
It's now about how much you can trust agents and how many work threads you can manage at a time before your brain overloads. It's like managing a team of 100+ mid-level engineers where you have to review all their code while making sure they don't break prod. At some point, you become the blocker.
Yes, I am the bottleneck now.
I'm sure there's a better way to do this without reviewing every line of code, but I can't help myself.
Cursor, Codex, and others are experimenting with different ways of managing swarms of agents without losing your mind, but I don't think we are there yet.
We're still early.
Lately, my workflow changes every couple of weeks anyway, and I'm sure it's about to change again soon.
While I wait for that to happen, I'll try not to lose my mind keeping track of all these agents.
Cheers, see you on the next rant.