Shilong Liu, an AI researcher active on X, published a taxonomy this month that sorts self-evolving agents into three distinct categories based on where the evolution actually happens. The framework names artifact optimization, harness self-improvement, and model learning as separate mechanisms, and it argues that conflating them is why “self-improving agent” has become a phrase that means almost nothing on its own.
The taxonomy matters because the term gets applied to wildly different systems. A chatbot that rewrites its own system prompt after a bad answer and a lab running continuous reinforcement learning on model weights both get called self-improving, even though one changes a text file and the other changes billions of parameters. Liu’s framework gives builders a sharper question to ask before accepting any such claim: which layer is actually evolving.
Artifact optimization is the layer most teams already touch without naming it. It covers agents that rewrite their own prompts, refine few-shot examples, or generate new tool definitions based on what worked last run. DSPy-style prompt compilers and self-critique loops that edit their own instructions both live here. Nothing about the underlying model changes; the artifacts surrounding it do.
Harness self-improvement sits one layer deeper. This is evolution in the scaffolding: an agent that edits its own orchestration logic, restructures its memory retrieval, or adds a new skill to a persistent library it can call later. Voyager’s Minecraft skill library and agents that rewrite their own subagent routing fall into this category. The model stays frozen. The system built around it gets smarter.
Model learning is the layer most people picture when they hear “self-improving AI,” and it is also the rarest in practice. This is genuine weight change: online fine-tuning, reinforcement learning from the agent’s own trajectories, or distillation back into the base model. It requires training infrastructure that most teams building on top of a hosted API simply do not have.
The distinction lands directly on the recursive self-improvement conversation that has circulated since GPT-4’s release. Most publicly demonstrated “self-improving” agents operate at the artifact or harness level, which is reversible and cheap to iterate on. Claims that imply model-level learning, the layer that would actually compound capability over time, deserve more scrutiny about what specifically is being updated and how.
For builders, the practical guidance is to invest where the taxonomy shows the highest return relative to cost. Artifact optimization is nearly free to experiment with and should be a default part of any agent’s feedback loop. Harness self-improvement, giving an agent the ability to extend its own tool library or restructure its memory, is where most of the real capability gains over the next year will come from, since it requires no retraining infrastructure at all. Model learning remains the domain of labs with the compute and data pipelines to run it safely, and teams without that infrastructure should stop describing prompt tweaks as if they belonged in the same category.
Based on a taxonomy proposed by researcher Shilong Liu on X in July 2026.
