Survey Redefines 'Self-Improving Agents': Fast Exploration, Slow Consolidation
新智元 · wechat · 2026-08-10
A joint survey from Jilin University, KAUST, and other institutions redefines "Self-Improving Agents," mapping model training, memory evolution, tool creation, and architecture search into a unified framework, alongside 312 related works.
The Boundary: True self-improvement occurs only when a system persistently modifies its foundation model parameters or scaffolding (Prompts, Memory, Tools) based on execution trajectories, altering the starting point for future tasks. One-off reflections don't count.
Two Core Paths:
- Foundation Model Improvement: Updates parameters via self-generated data or RL. Powerful but costly, with risks of bias amplification.
- Scaffolding Improvement: Rewrites Prompts, manages Memory, or creates Tools without changing model weights. Lightweight and reversible, making it the current engineering favorite.
Design Principle: Mature systems should adopt a dual-timescale approach: using scaffolding for fast, reversible exploration, and distilling proven experiences into model weights for slow consolidation. As agents gain self-modification capabilities, robust safety mechanisms like layered gating are crucial to prevent persistent vulnerabilities.
More from coding & agent
- Open Source iCloud MCP: Lets Claude Control Native Apple Apps — Glittering_Rule_6039 · 2026-08-10
- PyUIBuilder: Visually Build Python GUIs Like Canva, Hits 2.5k Stars — tom_doerr · 2026-08-10
- Garry Tan Shares Keynote on Deploying Real Agentic AI — garrytan · 2026-08-10
- llm-wiki: Open-Source Knowledge Base Compilation for Multi-Agent Research — tom_doerr · 2026-08-10
- BlackBox-MCP: Orchestrating Multiple AI Coding Agents — TheNightPorter28 · 2026-08-10
- Security Through Obscurity is Dead: AI Agent Swarms Will Exploit the Internet — nptacek · 2026-08-10