Why Prompt-Only Agents Break at Scale

A prompt-only agent has no cycle, no verification, and no weight update. It cannot adapt from its own outcomes.

What does a prompt-only agent lack?

The cycle, the verification step, and the strategy weights. Behavior comes from the prompt alone.

Can it be extended?

Yes, but extension adds components. It does not change the fact that the cycle belongs to the kernel.

When is prompt-only acceptable?

When the task has a fixed, single-step shape and no need to adapt.

Why does it break at scale?

Because it cannot adapt from its own outcomes. Without verification and weight updates, it cannot learn from its actions.

What is the difference between a prompt-only agent and a kernel?

The kernel runs the cycle itself and updates its own weights. A prompt-only agent relies on the prompt alone.