Closing the Consistency Gap: Self-Evolving Agents That Learn to Stay on Course

TL;DR - This paper introduces a self-evolving agent framework that learns from unstable trajectory steps to make repeated executions more consistent. On AppWorld, it substantially improves the share of tasks that succeed across all five runs.

• Defines the “consistency gap”: ReAct with GPT-4.1 averages a 77% per-run pass rate, but succeeds in all five attempts on only 53% of tasks.
• A Consistency Analyzer identifies steps likely to produce divergent outcomes across executions.
• A Guideline Generator converts these diagnoses into targeted episodic memories for future runs on related tasks.
• Five-run success rises by 16 percentage points on same-task evaluation and 13 points on similar-task generalization.