# How this website’s AI demos work: architecture and limitations > These pages are evaluation demos built around shared agent definitions. A successful model request can produce fresh output; a recorded fallback is an example, not an analysis of new input. The reviewed code does not establish that every named agent is an always-on production system or a ready-to-purchase product. Canonical page: https://withhammad.com/blog/71-live-ai-agents-website-architecture-lessons Published: 2026-08-09 · Updated: 2026-09-20 This article describes local code reviewed on 20 September 2026. It is not a verification of the production deployment or customer account integrations. The earlier article described dozens of live agents and repeated business results without establishing their supporting evidence. Those assertions have been removed; the original URL remains so existing links continue to work. ## A catalogue of definitions, not a fleet of autonomous workers A named agent page gives a visitor a focused input form and an output presentation. Shared configuration and runtime code make those pages reusable. A catalogue entry alone does not prove that a background worker is running, a customer account is connected, or a complete automation can be downloaded. Judge each page by its disclosed inputs, available tools and evaluation scope, not by its name. ## What the shared engine implements The relevant implementation is src/lib/agent-engine/getAgentOutput.ts. It runs input through a PII-scrubbing helper, optionally fetches a page for configurations that explicitly declare a webFetch pre-tool, and passes the prepared input to the generation helper. PII scrubbing should not be treated as a guarantee that every sensitive value will be detected: use fictional or non-sensitive inputs in a public demo. For the optional fetch, the engine requires an HTTPS URL. If the fetch fails, generation can continue from the visitor’s input alone. Typing a URL into another field therefore does not establish that the page was read. A generated answer must not be mistaken for verified access to a website, CRM, advertising account or calendar. ## Generated output versus recorded output After generation, the engine checks for a non-empty array of recognized output blocks, such as tables, checklists and metric cards. When that check passes, it returns mode live with the output and provider information. When it does not pass, the engine returns the configured example blocks with mode replay. A replay demonstrates presentation; it does not analyze newly submitted information. Shape validation is not fact checking. A well-formed table can still contain wrong numbers or unsupported conclusions. Review generated work against its sources before using it in a campaign, customer conversation or business decision. The engine also returns a prompt version and can record tracing spans when a tracer is supplied; this does not establish comprehensive production monitoring or a service-level guarantee. ## What this walkthrough does not prove This code review does not verify scheduled outreach, autonomous ad changes, customer results, continuous availability or successful account integrations. It does not establish that every catalogue item is a production-ready product. Those claims would require separate deployment checks, authorized integration tests, failure handling and traceable evidence. Public demo availability is not a substitute for that work. ## How to evaluate a demo Start with a fictional brief. Check whether the displayed run is generated or recorded, whether a tool actually accessed the requested source, and whether the output answers the brief. Inspect factual statements and calculations yourself. Before commissioning a custom build, agree on required integrations, human approval steps, data handling, acceptance tests and ongoing support. These are requirements to establish, not promises implied by the demo page. ## FAQ ### Are all agent pages always-on production systems? No. Catalogue pages and shared definitions do not establish continuously running workers or connected customer systems. ### Does a replay analyze my new input? No. A replay is a configured example used when fresh output is unavailable or invalid. ### Does output validation verify facts? No. The engine checks output structure; factual accuracy and business suitability still need human review. --- Written by Hammad Yousuf — AI Marketing Automation Engineer. https://withhammad.com