What the Agents Can't Take
I keep telling people that anything repeatable gets eaten by agents eventually. They hear a horror movie. I mean it as a map.
A map needs an edge. A line where one thing ends and another starts. The agents have a clean one. If a task can be specified, shown a few thousand times, and graded, it’s on their side. Which, if I’m honest, is most of my job. What does a backend engineer actually do on a Tuesday? Wire the endpoint. Write the test. Reconcile the ledger. Follow the runbook. Repeatable, all of it. None of it safe. Pretending otherwise just means you get to be surprised later, which is the worse version.
So the question isn’t what survives. It’s what’s on my side of the line. The honest answer: the stuff nobody can specify, because nobody knows what it should be yet.
An agent will build whatever you describe. It won’t tell you the thing isn’t worth building. It optimizes the metric you hand it. It won’t mention that the metric is wrong, that you picked the easy one because the real one is hard to see. It’ll drive your plan off a cliff with beautiful uptime the whole way down. Taste, picking the problem, knowing which schlep is worth the suffering and which is just suffering. That’s not a skill the agents are temporarily short on. It’s a different job. The job of saying this, not that, before anyone can prove you right.
The part people skip: as execution goes cheap, judgment gets expensive. When building costs nothing, all the value runs to deciding what and why. For most of history those were stapled together. You couldn’t think about a bridge without burning years learning to build one, so the builders made the calls. The agents are unstapling them. The people who only had the building half are not gonna make it. The people who had the deciding half just got handed a workforce.
Great news for one kind of person. Terrible for the other. If your value was being a reliable executor of known tasks, the floor is rising to your chin. If your value was knowing which task mattered, you just got leverage no generation before you had. Paine wrote pamphlets one at a time and moved a continent. Give that man a printing press that also writes and see what happens.
I don’t fix the anxiety by getting faster at the repeatable parts. That’s racing the machine at the one event it trained for. I fix it by spending my time on the parts you can’t write down. Picking problems, building taste, owning the judgment that decides where the agents get pointed.
They’re coming for everything that can be written down. So stop racing to write things down faster. Be the one who decides what gets written. Or something like that.
One of a series of essays. I’m Prajjwal Chittori. prajjwalchittori.com.