New survey maps how AI agents self-improve with minimal human input

A comprehensive new survey by Zhe Ren, Yimeng Chen, and collaborators (including Jürgen Schmidhuber) tackles the rapidly maturing field of self-improving autonomous agents. Published on arXiv in July 2026, the 97-page paper (with 12 figures) proposes a unified system-level framework: a modern agent is a configuration that couples a foundation model with an operational scaffold of prompts, memory, tools, and control logic. Within this view, self-improvement is formalized as a self-induced update operator that obtains and commits changes to either the model’s parameters or the scaffold components. The goal is controllable evolution from experience with minimal or no human input.

The survey organizes prior work by update target (parameters vs. scaffold) and by the signals driving change (e.g., reward, self-generated data, environmental feedback). It then reviews applications spanning robotics, software engineering, and interactive systems, followed by a discussion of evaluation practices and current limitations. The authors highlight open problems such as safety guarantees, scalability of self-improvement loops, and the risk of reward hacking. For researchers and engineers building next-gen AI agents, this paper provides both a structured taxonomy and a roadmap for future work.