Your AI Model Got Smarter. Why Does Coding Feel Worse?

Every time a new AI model drops, the benchmark charts go vertical. And yet, when you actually sit down to code with it, the “wow, this is clearly better” feeling rarely keeps pace. Sometimes you catch yourself thinking the opposite: that older tool felt smoother. That gap between benchmark and vibe is the whole story here.

The Problem Isn’t the Model. It’s the Harness.

Start with the anatomy. An AI coding tool is really two things stacked together.

The first is the model — the brain that reasons and writes the code. The second is the harness: the plumbing that reads files for the model, runs its commands, and pipes the results back. The model thinks. The harness fetches, executes, and reports.

Picture a kitchen. You can hire the best chef alive, but hand them dull knives, hide the ingredients, and deliver the wrong plate to the wrong table, and dinner falls apart. That’s the fault line in the current debate. The chef keeps getting better. The kitchen doesn’t — and increasingly, it’s the kitchen tripping the chef.

This is roughly the point developers like Armin Ronacher have been hammering on. The sharper the model gets, the more the friction from the harness stands out.

Why a Better Model Makes Bad Tools More Annoying

Here’s the actual paradox. Common sense says a smarter model should mean a smoother experience end to end. In practice, you often get the reverse.

The reason is that the bottleneck moved. Back when models were genuinely weak, clunky tooling didn’t register — the model was going to be wrong half the time anyway, so who’s counting? As models got smart, the math flipped. The source of error migrated from the brain to the plumbing.

Concrete example. A smart model correctly decides it needs to fix line 300 of a file. But the harness only showed it a truncated slice, and line 300 was never in view. The model did nothing wrong. The tool blindfolded it. When the model was dumb, that mistake vanished into the noise. When the model is sharp, “how did it miss that?” jumps off the screen.

Context Management: The Quiet Trap

The place the harness meddles most is context management. A model can only look at so much at once. A codebase is far bigger than that window. So the harness decides, on the model’s behalf, what to show and what to cut.

The trouble is that this decision is cruder than you’d hope. When a conversation runs long, the harness summarizes or trims the early parts — and important details get thrown out in the process. You know the scenario: you clearly said something twenty messages ago, and now the model acts like it never heard it. The model didn’t forget. The harness erased the memory.

There’s a twist here. You’d assume bigger context windows fix this, but not quite. Give the harness more room and it tends to cram in more material, which buries the signal you actually needed under fresh noise. It’s a textbook case of tool design lagging behind model capability.

More Tools Don’t Make a Better Agent

Another trap: tool overload. Bolting a hundred capabilities onto an agent feels like it should make it more powerful. Usually it does the opposite. Too many options and the model gets confused, burning reasoning just to decide which tool fits the moment.

Give a person ten near-identically named commands and watch them hesitate every single time. Models are no different. A well-built harness isn’t the one with the most tools — it’s the one that hands over exactly the few that matter, clearly. Less really is more here.

So the mark of a good harness is restraint, not spectacle. Stay out of the model’s way, pass it only what it needs, and trust the model with the rest. Easy to say. Brutally hard to actually balance.

The Next Battle Is the Plumbing, Not the Brain

Pull it together and the picture is clear. The next front in AI coding probably isn’t raw model intelligence — it’s harness design that lets that intelligence flow through unimpeded. The best brain in the world doesn’t help if the pipes leak.

Which means judging a new tool by benchmark scores alone is half an answer at best. Ever felt that the model seems brilliant but the whole thing is somehow sluggish and frustrating? That frustration might not be the model at all. It might be the plumbing hiding right behind it.