From God Knows to Claude Knows

There was a version of this job, not very long ago, that ran on trust.

Picture a Technical Lead. An alert goes off at 2 AM. A system is down, or a bug has slipped into production. The CTO doesn’t know the fix. But here’s the thing, they don’t need to. They know

exactly who to call. They know that person, they know how that person thinks, and they know how to get the best out of them. And the person on the other end? They carry the full context of that system in their head. They know whether this is a quick patch, a refactor, or a ticking time bomb. They know the urgency without being told.

So the CTO goes back to sleep. Not because they reviewed a single line of code, but because they trust the system of people underneath them. Everyone in their own silo, doing what they’re supposed to do. That was the old world. Call it “God knows”, somebody, somewhere in the chain, actually knows.

Now we live in a different world. The world of agents and sub-agents. And I want to be careful here, because I’m not being a pessimist, and I’m not doubting what the model writes. I’m pointing at something quieter.

Today, everybody is a system architect.

Everybody understands everything, in and out of the system, and where they don’t, they know they’re just one prompt away from understanding it. Correct or wrong, doesn’t matter. The feeling of knowing is always available. And that feeling is comforting. Dangerously comforting.

The false comfort

This is the part that doesn’t get said enough. The output looks confident, so we feel confident. But the two are not the same thing.

There’s a study from METR that I can’t stop thinking about. They took sixteen experienced developers, working on codebases they knew well, and measured them with and without AI tools. The developers were sure AI made them faster, they estimated a 20% speedup. In reality, they were 19% slower. The gap between how it felt and what actually happened wasn’t a rounding error. It was a complete disconnect.

That’s the false comfort, measured.

And it doesn’t just live in your head. GitClear looked at 211 million lines of code. Copy-pasted code is climbing. Refactoring, the boring, unglamorous work of consolidating things into one well-named place, has fallen off a cliff, dropping below 10% of changed lines. Duplicate blocks shot up several times over in a single year. We are shipping more code that does less consolidating. More boxes, fewer finishing touches.

The carpenter and the box

Let me give you the analogy I keep coming back to.

Say I want a wooden box. If I hand a design to a carpenter and never talk to them again, I find out what I actually got only when it’s delivered. But if I sit with the carpenter, watch them work, ask questions, take inspiration midway, I can shape it. I have time inside the process to add value.

Now say there’s a machine that makes the box in five minutes. Be honest: in five minutes, what value am I adding? None. So the box comes out... fine. A normal box. It lacks the small touches that only show up when a human is in the loop. And worse, it’ll take me another ten or fifteen days of living with that box to notice what’s wrong with it, and then I iterate again. And again.

And here’s the trap. The machine has a very sweet tongue. It doesn’t just make the box. It gently suggests that my static-site box also needs an S3 bucket, a queue, a CloudFront distribution, a WAF, maybe a Redis layer for good measure. Suddenly my simple box has an observability problem I now have to babysit. And if anything goes wrong — well, Claude is your boss now. You’ll ask it to debug the very thing it talked you into.

The deployment math

Here’s the part the dashboards hide.

In the pre-AI era, maybe I shipped ten things. But those ten were reliable. Now my sample size is a hundred. Even if my hit rate stayed exactly the same, simple arithmetic says thirty or forty of those hundred are going to fail. More shipping is not more reliability. It’s more surface area for doubt, and more than your cognition can actually hold.

The industry numbers point the same way. Google’s DORA research found that as teams lean harder into AI, delivery stability tends to drop, by single digits, but consistently, year over year. The throughput goes up. The confidence in the output stays low. We are getting faster at producing things we trust less.

The PRD passes. The tests are green. The docs even look great. But “it passed what was asked” is not the same as “it is reliable.” Reliability was the actual expectation. The ship that sails out is supposed to stay afloat. And when it sinks midway, supply-chain attack, a dependency you never read, who knows, the situation is different every single time.

The 31-year-old running a 200,000-person company

This is what worries me most.

A 31-year-old engineer today, with enough agents under them, is sitting at the operational complexity of someone running a two-lakh-person multinational. They’re watching metrics, KPIs, ten products, fifteen engineers, twenty agents on each. On paper, they’re a delivery head.

But the fifty-year-old who actually runs that MNC paid their dues to get there. Two decades of choosing which signals matter, which to ignore, which KPI is real and which is noise. You cannot compress that into a prompt. Asking a 31-year-old to instantly carry a 50-year-old’s judgment isn’t ambitious. It’s not humanly possible.

So what do we do

I don’t think the answer is to stop using these tools. That ship has sailed too.

The answer is simpler and harder: ship it like you coded it by hand. Ship it like you know it in and out. Because if you haven’t paid your dues to the feature, to the product, to the company, if you’re just forwarding the model’s confidence as your own , you eventually slide into a kind of AI psychosis. “God knows” quietly becomes “Claude knows,” and you stop being the one who knows anything at all.

Which is why I’ve started believing the real leverage isn’t in better prompts about what to ship. It’s in stricter rules about what not to ship.

That’s where the judgment lives now. That’s the part no agent can hold for you.