Structured self‑evaluation now demonstrably lifts LLM agent success on long‑horizon games. SkillCoach’s rubric loop turns vague outcome checks into a quantifiable process, substantially improving evaluation quality and providing stronger supervision signals.
Before these advances, most autonomous agents were judged solely by final task success, and memory contracts simply concatenated raw transcripts, making it impossible to isolate the effect of individual skills or memories. AgenticSTS introduced typed retrieval with bounded prompts to enable clean ablations, but still relied on outcome‑only filtering for supervision.
Gold‑keypoint coverage climbs from 71.56 % to 83.70 %, proving that the self‑evolving rubrics become substantially more complete with respect to human‑gold process evidence [1].
Usability scores rise from 81.53 % to 94.33 %, indicating that evolved rubrics give clearer criteria and negative cases for trajectory judging [1].
Hallucination rate drops from 2.00 % to 0.00 %, showing that the rubric expansion does not introduce unsupported constraints [1].
Baseline‑strict runs achieve a mean score of 70.4, confirming the task sits in a hard but non‑saturated regime [2].
When triggered strategic skills are enabled, win probability doubles from 3 / 10 to 6 / 10 games [2].
The reported Fisher exact p≈0.37 indicates the observed win‑rate increase is not statistically decisive, so further trials are needed to confirm robustness [2].
Future agent development pipelines should replace pure outcome filters with self‑evolving rubric supervision as the default evaluation layer, because richer process feedback translates into measurable performance gains.
