A marketing agent is just three things: code, a thinking loop, and a live data stream

The architecture of a marketing agent breaks down into three components. First, code — typically a cron job or plain scripts; the advice is to skip heavy frameworks because they are usually just bloat. Second, a thinking loop — an LLM call used only where judgment is genuinely needed, such as assessing ICP fit or writing reply copy. Third, a live data stream — new posts, engager lists, webhook replies, and analytics flowing back in. The guiding rule is: only use inference when you actually need it. Everything else is just scheduled automation.