What's this about?
A survey of self-improving agentic systems — AI agents that adapt from their own experience with minimal human input — arguing they've moved from research prototypes to deployed systems, and framing the field around "controllable evolution": getting the improvement without losing the reins.
What exactly is a "self-improving agent"?
An agent that gets better *while deployed*: caching successful strategies as reusable skills, rewriting its own prompts and tools, learning from failures — without a human retraining it. The survey's framing term, controllable evolution, names the tension: adaptation is the point, but unsupervised drift is the risk (echoing Anthropic's misalignment findings covered yesterday).
Why does it matter?
Surveys like this mark a field crossing from papers to practice — a map of techniques that agent products (including coding agents) are already quietly shipping. (HF 14↑)