Do self-improving agents really split into two distinct loops?

A survey of modern self-improving agents offers a clean systems abstraction: an agent is a configuration coupling a foundation model with an operational scaffold — prompts, memory, tools, and control logic — and self-improvement is a self-induced update operator that obtains and commits updates to either the model parameters or the scaffold components. That single distinction organizes the whole field, because it separates what changes from how fast it can change. Foundation-model improvement is the parametric, slower loop (driven by intrinsic generative demonstrations, intrinsic evaluative feedback, or extrinsic exploratory experience). Scaffolding improvement is the non-parametric, faster loop (updating prompts, memory, tools, or the full scaffold).

The value of the framing is that it makes the design space legible rather than a pile of named systems. It also explains why so much recent progress lives in the fast loop: updating a memory store or a prompt is cheap and reversible, while updating weights is expensive and risky. This is the same two-timescale structure surfaced empirically in Can agents adapt without pausing service to users? — but generalized from one system to a taxonomy. Which means the survey's real contribution is a shift of the object of study: the goal is no longer the static agent but the mechanism driving its evolution — the update operator itself and the signals that trigger it. Any concrete self-improvement design can be located by asking which loop it modifies and what signal drives the commit.