I got the Claude Certified Architect Foundations credential this summer and the biggest lesson from passing the exam wasn't how to build an AI agent. 👉 It was when you shouldn't bother. The Exam puts the Biggest Weight on the Agentic Orchestration - Domain 1 & For Very Good Reason 👀 So here's what I wish all the AI Hype men on YouTube would have told me about agents. First, what is an "agent"? Think of a new employee you can give a goal to, not a checklist. You say "sort out this unhappy customer," and they figure out the steps: look up the order, check your policy, decide what to do. That's different from automation, which will complete the same process every time. 1. If the steps are always the same, like sending a receipt after every purchase, simple automation does it cheaper and more reliably. 2. Agents are for work that's different every time. Agents are a good fit when: - Every case looks a little different - there is open ended creativity allowed as agents are probabilistic - The task involves checking several places (email, spreadsheets, your CRM) - The information is messy, like long customer emails or scanned invoices - A person can quickly review the result before it matters 3. Telling AI to do something isn't a rule. You can tell an AI "never give refunds over $500," and it will usually listen - 95% of the time? Usually isn't good enough for a business. The real fix is building the guardrails in your system so the AI can't go past it, the same way a cashier can't refund more than the register allows. 4. Knowing when to escalate to a human is a necessary feature. The best agents know what they shouldn't decide on their own. Angry customer? Big refund? Legal question? It hands off to you with the whole story already written up. Here's what an agent that works in a real business can look like: - The sales call prep. Before every call, the agent reads the prospect's website, recent news, and your past emails with them, then hands you a one-page cheat sheet. No more 45 minutes of tab-hopping.