Source: arXiv — 2026-07-14
Summary
A new survey, co-authored by Jürgen Schmidhuber, formalizes how modern agentic systems self-improve once deployed, framing today's typical agent as "a foundation model plus an operational scaffold" and taxonomizing the mechanisms by which that combination gets better over time without full retraining. The paper is notable both for its author pedigree in self-improving-AI research and its timing, as self-improving agents move from research demos toward production deployments.
Key Takeaways
- The survey's central framing splits an agent into two separable layers — the underlying foundation model, and the "operational scaffold" (harness, tools, memory, prompting strategy) wrapped around it — and argues self-improvement often happens at the scaffold layer, not by retraining the model itself.
- This maps directly onto real-world practice: teams that tune prompts, tool definitions, and memory strategies around a fixed model are, in this framing, doing a lightweight form of self-improvement without touching weights.
- Schmidhuber's involvement gives the survey extra weight given his long research history specifically on self-referential and self-improving AI systems, predating the current agent boom by decades.
- As a survey rather than a benchmark paper, its value is in the taxonomy and framing for practitioners trying to reason about where self-improvement is actually happening in their own agent stack, rather than in a specific new technique or number.