I've been building a small autonomous outreach agent as a portfolio case study: https://lnkd.in/ewygDbuS The idea was not to build a “fully autonomous AI salesperson”. I’m more interested in the practical version of agents: systems that do the repetitive operating work in the background, but still keep a human in control when something external-facing or reputational is involved. This one imports broker records, drafts personalised emails with Claude, queues them for approval, sends from a dedicated inbox, schedules follow-ups, reads replies, classifies them, escalates the useful ones and produces a weekly report. The part I care about most is the architecture. The workflow logic handles the things that need to be reliable: status, timing, retries, limits, approvals, suppression rules and reporting.The LLM handles the judgement-heavy parts: drafting, classification and suggested replies.Everything important is logged. Nothing sensitive goes out without approval. What was useful is that the build caught real issues during testing: One was a silent database write failure, where the system reported success but had not actually written the record. Another was a send guard that was too strict and blocked an approved reply. That is exactly why I think these systems need proper workflow design, not just prompts. For me, the useful pattern is: Not a chatbot. Not a black box. A backend workflow system with an LLM inside it, an approval layer around it, and a dashboard for visibility, exceptions and trust.