The durable value is moving below the model

Two years building autonomous agents in one domain taught me something I'll state with the caveat it deserves: this is n=1. One engine, one field, one team. Treat what follows as a hypothesis generator, not a settled finding. The observation that would change my mind is easy to name: show me a domain where a better base model alone closed a gap that took a team years to build underneath it, and I'll update.

With that bound named: the thing I keep seeing is that the model is rarely the binding constraint.

The engine is NovoMCP, computational chemistry for drug discovery and materials. The reasoning layer is good and getting better on a schedule I don't control. What breaks is everything around it. Notes from the floor, not forecasts.

The wall is the long-running work, not the harness

When you start, you worry about harness engineering, about squeezing the model. That worry is real and mostly tractable. It converges. The wall shows up later, when the work has to actually run.

A free-energy calculation in our pipeline runs three to four hours. Making it durable, resumable, and cancellable (so a dropped connection doesn't kill a job you've already paid for in GPU-hours) was harder than any prompt we ever wrote. None of that engineering is visible to the user, and none of it is about intelligence. First-order, the difficulty of an agent system tracks the runtime of its longest unit of work, not the cleverness of its reasoning.

Coupling happens at the reasoning layer. The domain layer doesn't move.

There's a view going around that the era of one generic harness with swappable models underneath is ending, that the next generation couples the harness tightly to a single model. In the regime I can see, that's true, and it matters less than it sounds.

The coupling lives at the reasoning layer. We've swapped the model driving our engine more than once. Each time, the reasoning got sharper. The 122 million compounds we'd already characterized did not change. The compliance checks did not change. The record of what the system tried and rejected did not change. The expensive layer is precisely the one that doesn't care which model reasons over it.

"Did it work" has an answer here, and that's a special regime, not a universal one

A lot of agent work can't define success, so it falls back on evals. Chemistry is kinder, and I want to be honest about why: not because we're smarter, but because the domain hands us verifiable outcomes for free. A compliance check passes or it doesn't. A candidate clears a gate or gets filtered. A docking score is a number, wrong sometimes, but a number you can check against ground truth.

When the outcome is verifiable, you can let a system run further on its own, because you trust the gate to stop it. That's a structural advantage worth building around. It's also not portable. The general agent world is working hard to manufacture verifiable outcomes; some domains were born with them. Modulo one caveat I'd hold myself to: a gate you can't validate is just an eval wearing a lab coat. Verifiability you haven't checked against ground truth doesn't count.

As reasoning gets cheaper, the value moves down, first-order

The direction everyone's pointing at: hand a system an outcome and a budget, let it figure out the rest, including which model to use and how to wire itself together. I think that's roughly right, first-order.

The consequence people skip past: as the reasoning layer automates itself, the differentiated value moves below it, to the things a better model cannot conjure. You can't prompt a model into a hundred million compounds with their toxicity already computed, or a compliance method validated against a real benchmark, or an audit trail of every decision a run has ever made. That's earned, not generated. It takes years, and it doesn't photograph well in a demo.

Here's the way that claim could be wrong, so you can watch for it as well as I can: if synthetic data and self-play get cheap enough that the domain layer becomes generatable on demand, the moat I'm describing evaporates. I don't think we're there. But I'd want to see it converge before I bet against it.

Where this leaves me

This isn't an argument to rebuild the unglamorous layer yourself when someone has already built it. It's an argument to be clear-eyed about where the durable value sits. For a long time the smart money was on the model. My read (n=1, bounded above, falsifiable on the terms I gave) is that the next few years quietly move it underneath.

---

We build that layer for chemistry. If you work where AI reasoning meets a real scientific domain, I'd be glad to compare notes, especially if your experience cuts against mine.

— Ari Harrison