When AI Agents Change Without Getting Better

An AI agent that learns from experience may leave behind written reflections, executable skills or revised code. Those artifacts offer evidence that something has changed. But do they show that the system has become more capable? In the arXiv preprint Artifact-to-Capability, Ziqian Ban and colleagues argue that accumulating artifacts, changing behavior and raising selected metrics are not enough to answer that question.

The researchers examine implementations of three self-evolving approaches—Reflexion, Voyager and DGM—in freestyle Gomoku. Played on a 15-by-15 board, the game rewards completing a line of five stones. Its value here is experimental control: researchers can restore exactly the same position, check moves against independent rules and inspect the sequence of decisions that produced an outcome.

All three implementations used the same base model, DeepSeek-V4-Flash, and the same programmatic opponents during evolution. The researchers then froze the resulting artifacts for diagnosis. These are particular implementations in a particular environment, not comprehensive evaluations of everything the three approaches can do.

What should count as a capability?

The authors propose five requirements. An artifact must exist and enter decision-making; it must change behavior beyond ordinary sampling variation; those changes must improve task-relevant decisions; other components must not override an established useful effect; and later versions should extend useful behavior while preserving what already works. Under their definition, the middle two requirements form the checkpoint for capability formation. Extension is a separate question about subsequent evolution.

They examine these requirements at three scales: individual board positions, complete games and successive code versions. The fixed-position evaluation contains 1,876 states, combining constructed tactical positions with positions collected during DGM evolution. A rule-based checker identifies acceptable moves. Importantly, the authors distinguish making a legal move from choosing a useful one: avoiding an occupied intersection does not, by itself, demonstrate better tactical selection.

Three different breaks

Reflexion stored textual lessons and included them in later move prompts. Loading those memories changed the model’s choices beyond the variation observed between two memory-free runs. Overall pass rates also increased. Yet the proportion of legal moves that passed the tactical check improved in none of the 12 memory-and-state-set combinations tested.

The authors call this a “legalization gain.” Their diagnosis attributes the overall improvement to fewer illegal moves, rather than better selection among legal alternatives. On the larger state set, additional paired analysis of positions where both conditions produced legal moves also showed a negative net change for the two memory conditions examined. This does not make legality improvements worthless; it means they do not satisfy the paper’s separate criterion for tactical capability.

Voyager exposed a different problem. Its evolved skill library improved fixed-position performance relative to the first-round library. In complete games, however, the system failed to take an immediate win in all 20 games tested and lost each game.

The recorded call order explains the failure. Whenever both players had an immediate winning opportunity, the calling script checked the opponent’s threat first, blocked it and returned before checking its own winning move. Completing its own line would have ended the game immediately. The authors describe this as a “compositional override”: useful component behavior was defeated by the way components were combined. The 20 games followed four distinct move sequences, so they should not be read as 20 different kinds of tactical failure.

DGM did pass the authors’ capability checkpoint at version OB-3. It improved fixed-state tactical performance relative to the initial strategy and took every immediate winning opportunity encountered in its diagnostic games. That finding did not mean it surpassed the initial strategy on every measure: OB-3 won 19 of 20 games against the strongest opponent, while the initial strategy won all 20.

Later versions failed to extend that capability. Two successive versions lost passing states without adding any on the larger fixed set. Wins eventually fell to zero, with some games ending undecided rather than in defeat. The final revision changed only one move across a separate pool of 7,234 positions. A composite evolutionary score nevertheless rose during one earlier regression because its game-length reward offset fewer wins.

The paper’s conclusion is narrower than saying self-evolving agents cannot improve. Its examples show why selected improvement signals cannot establish dependable capability on their own. The findings remain preprint results confined to Gomoku, these implementations and the reported runs. Different models, longer evolution and repeated runs remain untested; applying the framework to open-ended tasks without clear rule-based checks is also unresolved.