The Harness Gets Cheap. The Grade Doesn't.

Every AI agent has two parts. The model does the reasoning. The harness is the code around it. It decides what the model sees, which tools it calls, what happens when something fails.

If the model is the engine, the harness is the rest of the car.

Until recently, engineers built the harness by hand. Most of the craft in building agents lives there.

Two weeks ago a paper showed a model learning to rewrite it. This morning that got summarized as "the model swallows the harness," with Tesla FSD as the comparison.

I read the paper. The result is big, and the precise version is quite useful.

A second model rewrites the harness, runs it, checks the score, and rewrites again. Nothing inside the original model changes. The skill carries over to tasks it never practiced on.

The harness becomes something the model writes.

FSD was a different event. There the hand-written code was absorbed into the network itself.

If you're deciding where to spend, that difference is critical.

The harness is getting cheap to produce. Custom agent workflows will be rewritten, possibly by the model itself. Budget them like consumables.

The score doesn't get rewritten. Every revision in that paper climbs toward a grade on practice questions with known answers. Remove the grade and there's nothing to climb. Researchers are trying to let systems grade themselves. Whether those grades can be trusted is still the open problem.

In code, the grade comes free. Tests pass or fail.

In finance, legal, operations, it mostly lives in a senior person's head. What a clean reconciliation looks like. When a contract clause is acceptable. Rarely written down in a form a system can score.

That's where improvement stalls. The models get better every quarter. Your function improves only as fast as someone can score its output.

Defining what "correct" means takes domain judgment. That makes it leadership work. It belongs to the people who run the function.

Judgment was already the scarce input. Now the system needs it in writing.

AI will produce faster than anyone can check it. That stack lands on the desk of whoever knows what good looks like.