Right-Sizing LLM-Agent Decomposition in VAT Determination: A Pilot Controlled Sweep

TL;DR - A controlled 4,400-run pilot tests how finely LLM-agent workflows should be decomposed for cross-border VAT determination. Intermediate decompositions achieved the highest accuracy, but the preregistered evidence threshold was not met, leaving their advantage unconfirmed at pilot scale.

• Intermediate configurations reached 0.830 accuracy, versus 0.720 and 0.770 for the widest and most fragmented endpoints.
• A token-matched single agent trailed the leader by 6.5 percentage points, but the confidence interval included zero, so prompt budget may explain the difference.
• All configurations tolerated availability faults, while a schema-valid hallucinated record degraded every setup and harmed fragmented configurations most.
• The authors release the oracle, dataset, evaluation harness, raw traces, and analysis pipeline alongside a heuristic that partitions work near the dependency-layer midpoint.