New survey maps how agentic systems are learning to improve themselves

SchmidhuberAI · x · 2026-07-21

What the survey covers

The post introduces a new survey, “Self-Improvements in Modern Agentic Systems,” which reviews how classic ideas like meta-learning and recursive self-improvement are being revived by foundation models.

It links to the paper, project page, and GitHub repo, positioning the work as a broad look at how agentic systems are evolving through:

The accompanying figure maps a 2023–2026 landscape of methods across FM improvement and scaffolding improvement, showing a fast-expanding research area around memory, tools, prompting, and agent design.

Original post →

More from Embodied

Embodied channel →