Alibaba’s SPAR Lifts 1688 Conversion 15.8%

SPAR’s important contribution is turning agent self-improvement into an attribution problem: successful outcomes should update strategy only when the agent caused them. The results are compelling, but remain concentrated in Alibaba’s B2B sales domain and company-specific deployment data.

• –Inner-loop memory tracks each merchant’s objections, context, and follow-up strategy across sessions.
• –The outer loop promotes, refines, merges, or retires reusable sales hypotheses.
• –Runtime strategy evolution offers a practical alternative to fine-tuning for specialized agents.
• –The 200,000-plus merchant A/B test is stronger evidence than benchmark-only evaluation.
• –The core engineering lesson: store attributed outcomes and failure patterns, not just retrieved transcripts.