Most Copilot Studio talks show the build, the demo, the applause. This one shows the graveyard. Five real agents from real Jumpstart engagements that shipped, struggled, and got pulled. No vendor names, no client names, just the patterns that keep killing internal AI projects. The knowledge soup: an agent grounded on an entire SharePoint site that confidently returned policy from 2019 alongside policy from 2025 - why broad RAG without hygiene is a hallucination factory and what to scope instead. The over-trusting agent: a procurement assistant that took the user's word on their approval level and called the back-end API anyway - how agents inherit privilege confusion from the prompt and the simple identity check that fixes it. The expensive intern: an agent that worked beautifully and burned through tokens and Copilot Studio messages faster than the team it was meant to assist - how to instrument cost from day one and what to do when the curve goes vertical. The forgotten agent: built in two weeks, launched on a Friday, dead by month two - why 'we'll add adoption later' never works and the three things that make agents actually used. The frankenstein: sixteen Power Automate flows, four custom connectors, three knowledge sources - one upstream API change took the whole thing down, and how to design for the failure mode you will absolutely have. You leave with an honest mental model of what breaks, a 5-point pre-deployment checklist, and the confidence to say no when an agent is not ready to ship.