Case Study: Design an Ask-the-Web Agent

"Design a Perplexity-style agent that answers a question by searching the live web" is a design prompt that looks like a RAG system with a different corpus, and treating it that way is the most common mistake candidates make. The corpus here isn't a curated, pre-indexed knowledge base you control — it's the open web: unbounded, adversarial, stale in ways you can't predict, and reachable only through a search API and a fetch step that both cost real money and real milliseconds on every single query. A weak answer wires a search call to an LLM and calls it done. A strong answer treats "ask the web" as a pipeline with a freshness decision, a fetch step that must run in parallel or the latency budget is unrecoverable, a citation-faithfulness contract with the user ("every claim in this answer is attached to a source that actually says it"), and a threat model where every fetched page is an attacker's chance to inject instructions into the synthesis step.

This subject is a model answer. It assumes you've read agent-design-patterns-workflows-to-autonomous-agents — this pipeline is a direct, worked application of that subject's named patterns, not a fresh invention of similar ideas under new names — and guardrails-and-prompt-injection-defense, whose defense-in-depth layers this subject applies to the specific threat of untrusted fetched web content rather than re-deriving. Where a number appears, it is an assumption to state, not a fact to memorize.