Growth Mostly Happens Outside the Weights

How does an agent get better?

The intuitive answer: train a stronger model. Let it grow in the weights — deeper reasoning, more knowledge. That path exists, but it's heavy, slow, and expensive, and most of the time you don't need it — because the vast majority of everyday growth happens outside the weights.

Growth outside the weights

A track record changing judgment is already growth outside the weights. Widen the frame and there's a whole field of it:

Write up a new skill, and the agent can do something it couldn't before. Improve a template, and its output gets steadier. Record a preference, and it never asks twice. Bank one failure, and it steers clear of the same trap next time. Distill a better way of working, and it ramps up faster on the next task. All of these make the agent better at its job while the model hasn't moved an inch — it's still the same candidate.

That's also how a person becomes a skilled worker: not a new brain, but a stock of experience to draw on and methods worn smooth by use. The bulk of an agent's growth works the same way.

Try outside the weights first

Hence a very practical ordering: when you want an agent to improve, first ask whether the problem can be solved outside the weights.

It usually can. Missing a capability? Check whether it's missing a skill. Unstable? Check whether it's missing a template or a memory. Keeps making the same mistake? Check whether that lesson ever got written down. All of these are far lighter and faster than fine-tuning, and none of them touch the model. Reaching for "a stronger model" or "fine-tuning" as the first move is often using the heaviest tool on a problem the lightest one would have solved.

Growth in the weights isn't off the table. It should be reserved for the gaps that genuinely can't be closed outside them. Keep the two kinds of growth distinct, and resources go where they count.