Taller Mountains

We got handed a once-in-a-generation chance to rethink how work actually works. To look at the decades of accumulated organizational scar tissue — the managers managing managers managing managers watching one person do something — and finally ask: what if we just… didn’t?

Instead, we’re using AI to build taller mountains.

Same bureaucracy. Faster output. Mountains of markdown generated by people who won’t read it, summarized by AI for people who won’t read the summary, filed into systems that exist for the sole purpose of proving that a process was followed. Nobody’s checking what they’re writing. Nobody’s asking whether it should be written at all. The box got checked. That’s what matters now.

The gatekeeper class found a new currency, and it’s text vomit.

All productivity is not good productivity. Spewing mountains of generated documentation doesn’t make you an AI-powered organization — it makes you the same organization with a faster printer. The disease was never that the reports took too long to write. The disease was that the reports existed in the place of the work itself. AI didn’t cure that. AI gave the disease a turbocharger.

And here’s what should really bother you — the same pattern shows up in how we build AI tools themselves.

I watch teams architect elaborate multi-agent, multi-step workflows before they’ve spent a single afternoon just talking to the model. They’re designing the recursive factory pattern — agents spawning agents, orchestration layers, routing logic, fallback chains — for a problem they haven’t manually solved once. They’re engineering a cathedral when they haven’t even confirmed the cathedral goes here.

It’s premature optimization. It’s always been premature optimization. We just gave it a new costume.

Before you build the fancy agentic workflow, do something radical: just use the AI directly. Hold down the speak button. Talk to it. Do the work manually, with AI, over and over. Through that repetition — not through architecture diagrams — you’ll feel where the friction actually is. Not where you imagine it’ll be. Not where a blog post told you it’d be. Where it actually shows up, in your hands, doing real work.

Then — and only then — commit to structure. A simple Python script, even one the AI wrote for you, can be a better mechanism for automation than a sophisticated multi-agent pipeline. Because the script solves a problem you confirmed exists. The pipeline solves a problem you hypothesized about in a meeting.

Assuming a structure before you’ve felt the pattern is how you end up with the software equivalent of optional flags on every function and nullable fields on every model. Technically flexible. Practically unmaintainable. You built for every possible future except the one that actually showed up.

The instinct to systematize is a good instinct. But it has a timing problem. Structure applied too early calcifies around assumptions. Structure applied after repetition calcifies around reality. One of those holds up. The other becomes the thing the next team has to work around.

This isn’t an anti-AI take. I’m building my entire company on AI. I’ve trained tons of engineers on AI-augmented development. I’m as bullish on these tools as anyone alive. And that’s exactly why this matters — because the people most capable of using AI to flatten bureaucracy and eliminate waste are instead using it to generate more artifacts for the bureaucracy to process. The most powerful simplification tool ever built, conscripted into the service of complexity.

The move is always the same. Do the simple thing first. Talk to the AI. Repeat yourself. Notice the patterns with your hands, not your whiteboard. And when you finally feel the friction — the real friction, the kind you can point to because you’ve hit it fifteen times — then build.

Anything else is just taller mountains.