The Big Picture
Splitting the VAT-determination task into a few focused AI workers can improve accuracy (peaking at 83%), but gains come with higher token cost and slower responses—and some gains may be explained by extra prompt budget rather than agent count alone.
ON THIS PAGE
The Evidence
A middle-ground decomposition (neither one huge agent nor many tiny agents) produced the highest accuracy (0.830) on the VAT testbed. Accuracy was non-monotonic across designs: the wide single-orchestrated worker scored 0.720, the finest-grained setup scored 0.770, and the two intermediate splits both scored 0.830. Moving to more workers raised median response time and token cost; matched-token tests suggest some accuracy edges may be due to extra prompt budget. Under failure tests, availability faults were absorbed across designs, but a single schema-valid but incorrect record (a content hallucination) damaged fragmented setups most and could flip the ranking. Sub-Agent Delegation Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Accuracy by configuration: wide = 72.0%, intermediate1 = 83.0%, intermediate2 = 83.0%, fine = 77.0%.
2Best orchestrated gain over a tuned single-agent: +7.5 percentage points accuracy, at +14% token cost and +42% median latency.
3Failure effects: schema-conforming hallucination dropped fragmented configurations by ~46–48 percentage points vs the wide config drop of ~31.5 points; wide-scope restart recovered ~+16 percentage points in availability tests.
What This Means
Engineers designing production agent pipelines and technical leaders deciding whether to decompose work across multiple agents should care because decomposition affects accuracy, cost, and recoverability in tangible ways. Compliance and reliability teams running legal- or finance-critical flows should pay attention to how handoffs propagate content errors and plan validation or rollback policies accordingly. Dynamic Task Routing Pattern
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
This is a pilot on a bounded VAT-style task with 40 cases and one base model family, so results may not generalize to very different domains or larger corpora. The tuned single-agent baseline narrowed the gap in a matched-token test, meaning some observed advantages could stem from extra prompt budget rather than agent count alone. Failure injections were synthetic and limited—real-world correlated or adversarial faults could behave differently. Reasoning
Methodology & More
A controlled pilot varied only how subtasks in cross-border VAT determination were assigned to workers while keeping the task surface, tools, prompts, and merge policy fixed. Five conditions were evaluated: four orchestrated decompositions from wide to fine scope and a tuned single-agent baseline without an orchestrator, across 40 oracle-labeled cases with five repeats. Key metrics were final-answer accuracy, per-step accuracy, token cost, and median latency; the study also included a matched-token arm and three failure-injection modes (worker timeout, schema-conforming hallucination, shared-tool outage). Results show a non-monotonic relationship between decomposition and end-to-end accuracy. Two intermediate decompositions tied for the best accuracy (83.0%), beating a wide single-orchestrated worker (72.0%) and the finest fragmentation (77.0%), but they cost more tokens and increased median latency. A matched-token comparison with the tuned single-agent baseline did not conclusively rule out prompt-budget effects, so some of the advantage may come from richer role prompts and scratchpads. In failure tests, availability faults (timeouts, tool outage) were absorbed across configurations, while a single validated-but-wrong record propagated through handoffs and hurt fragmented setups most. Practical takeaways: boundary placement (which subtasks are grouped) matters more than sheer worker count, record validation is critical for safety-sensitive workflows, and validation-filtered resampling (wide-scope restart) can improve reliability at a measured token cost. Emergence-Aware Monitoring Pattern. Mutual Verification Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Single author with no affiliation or citation information and arXiv preprint — unknown/low credibility signal.