A business analyst builds a custom agent to summarize competitor pricing from official product pages. The summaries must reflect current pages, cite sources, and avoid stale training knowledge. Static instructions alone keep returning outdated prices. Which custom agent design with web search best fixes the currency problem?

The correct design succeeds because it replaces reliance on static instructions and parametric memory with an enforceable retrieval workflow that mandates a fresh search over the official product pages for every request, requires specific citations so each price can be traced, and defines explicit fallback behavior when pages are unavailable to prevent silent substitution of outdated knowledge. The cached-results design fails on currency because even recent snapshots can preserve superseded prices and promotions, which directly conflicts with the requirement that summaries reflect current pages and explains why speed optimizations cannot override live verification here. The aggregators-inclusive design fails on source authority because comparison sites and marketplaces may lag, aggregate variants, or introduce fees that differ from the official pages, so broader coverage comes at the cost of the required ground truth. The model-estimate design fails on knowledge control because allowing the model to fill gaps from training reintroduces the exact staleness problem the agent is meant to fix and undermines citation integrity when no live source exists. An exam caveat is that plausible efficiency or completeness enhancements are still wrong when they compromise a stated constraint, so prioritize currency, official sourcing, and no-training-data use over coverage or readability. As a practical check, change a test product price, rerun the agent, and confirm the summary updates, cites the opened official URL, and declares unavailability rather than caching, aggregating, or estimating.