What a 40-query benchmark taught me about building AI agents for the web

A short field report on retrieval, cost, coverage, and the traps in between. I spent a few days building a small multi-agent pipeline — Amazon Bedrock (Claude Sonnet) as the model, Strands for orchest