Three-Tier Framework for Self-Evolving Agents

青稞AI · wechat · 2026-07-12

This article introduces a classification framework for Self-evolving Agents, aiming to clarify frequently conflated concepts: the exact differences between self-evolving, self-improving, continual learning, and test-time training.

The article proposes three key layers: Models / Harness / Artifacts. Models handle reasoning; the Harness includes agentic peripheral components like loops, memory, tools, and skills; Artifacts are the outputs generated by the agent, such as discovered algorithms, papers, or robotic policies. The author notes that many recent "self-evolving" systems are primarily iterating and optimizing Artifacts, typically seen in coding agents driving scientific and algorithmic discoveries.

The article then isolates "Agent self-improvement," discussing how systems can continuously strengthen via prompts, memory, tools, and skills without updating model weights, further exploring multi-agent routing, expert division of labor, and context management. Finally, it separates "model learning without golden answers," covering areas like self-training, self-play, RL, test-time training, and continual learning, emphasizing that such work fundamentally involves updating model parameters rather than just improving the Harness.

Original post →

More from coding & agent

coding & agent channel →