Five levels to measure how far an agent can modify itself

A Tencent Hunyuan study classifies self-evolving agents from L0 to L4 and proposes a framework for verifying whether their changes represent genuine improvements.

A team from Tencent Hunyuan and several Chinese universities has proposed a framework for mapping agents that can progressively modify their own behavior and internal structure. The study is not only concerned with determining what a system can change, but more importantly with what evidence is needed to claim that a given modification actually represents an improvement.

The taxonomy distinguishes five levels based on the depth of the modification. L0 concerns only the response or trajectory produced for a given task. L1 reaches the model or its trainable policy. L2 covers persistent components surrounding the model, such as prompts, skills, memory, workflows, or the harness. L3 changes the way future improvements are proposed, selected, or rolled back. Finally, L4 directly affects the criteria used to judge behavior and future modifications.

This hierarchy is neither a performance ranking nor a mandatory progression. It simply measures how deeply a change propagates through the system. Starting at L1, changes must be able to influence future tasks in order to be considered persistent. Under the authors’ definition, recursive self-improvement begins at L3, when the agent can modify the mechanism responsible for producing its own future updates. This boundary, however, is specific to the study and is not a universal definition across the field.

The core of the work focuses on the reliability of this evolution. The authors argue for a simple principle: an update should not be able to control the only evidence used to prove that it is better. Evaluation, testing, and the decision to accept or reject a modification should therefore remain outside the scope of what the agent is allowed to change. External audits can rely on held-out tests, executable checks, controlled experiments, formal methods, or human review.

The study also associates each level with characteristic risks, ranging from self-confirmation at L0 to criterion drift at L4, with model collapse, scaffold overfitting, and metric capture in between. The deeper an agent can modify its own improvement process, the more important it becomes for evaluation to remain independent and to assess effects over time.

The project also includes an open catalog of 549 works used in the survey and classified according to this taxonomy: 42 references at L0, 137 at L1, 257 at L2, 21 at L3, and 30 at L4, alongside additional work on field positioning, reliability, and future research directions.