Schmidhuber Team Maps the Self-Improving Agent Design Space

A 97-page survey from the group around Jürgen Schmidhuber, posted to arXiv this month, tries to do something the self-improving-agents literature has needed. It gives the field a shared vocabulary before every lab ships its own definition into production.

The move that matters is the split between what gets updated and what drives the update. The authors formalize a modern agent as a foundation model plus what they call an "operational scaffold of prompts, memory, tools, and control logic," and then define self-improvement as "a self-induced update operator that obtains and commits updates to model parameters or scaffold components." That gives you two axes to place any system on. Is the model itself being retrained, or is the scaffold around it being rewritten? And where is the change signal coming from: internal generation, internal evaluation, or external experience?

Why this matters if you are not writing the next agent paper: the scaffold-versus-parameters distinction lines up with a real operational choice. Rewriting prompts, memory schemas, or tool routing in flight is a different risk profile from letting an agent update model weights, and a shared taxonomy makes it easier to say which layer any given product is actually touching. The authors are explicit that the "primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input," and just as explicit that this is not solved. They flag "self-deception, reward tampering, and fatal errors" as failure modes when agents act on "limited resources and irreversible actions," and note that even the classic EURISKO system "depended heavily on the user serving as an external evaluation signal" rather than closing the loop autonomously.

The honest caveat is that this is a survey, not an empirical study. It does not tell you which shipped products today actually qualify, how much human oversight has been removed in current deployments, or what evaluation benchmarks the community should adopt next. The claim that these systems are "moving from research prototypes to deployed systems" is asserted in the abstract, not quantified there.

What is useful for anyone building agents right now is the reframe. If your roadmap has an "adaptive" or "self-improving" feature on it, the paper gives you a way to describe exactly which layer you are letting evolve and which signal you are trusting to drive it, which is the conversation worth having before the safety review, not after.