Confluent asked 4,625 IT leaders across 14 countries what actually blocks them from scaling AI agents. The answers came back almost tied. Skills gap and organisational readiness, 69 percent. LLM reliability and non-determinism, 68 percent. Data infrastructure and quality, 66 percent. Governance, risk and compliance, 65 percent. Four points separate first place from last. There is no single villain here. That is the finding, and it is more useful than any ranking would have been. Because if one blocker dominated, you could go and fix that one thing. Four blockers sitting four points apart tells you something different. It tells you the problem is not located in any one component. Here is the systems lesson. An agent is not a system. It is one component inside a system. And a component is only ever as good as what surrounds it. Notice that only one of those four blockers is about the model at all. The other three are about the organisation around the model. Skills. Data. Governance. Four things have to exist around an agent before it can do real work. Clean inputs. An agent cannot reason its way past a messy source of truth. If your data lives in six places and three of them disagree, the agent inherits the argument. A defined output. Where does the work land when it is finished? A card, a doc, a channel. If nobody decided this, the agent produces orphaned work that no one ever sees. A handoff. Someone or something has to accept that output and act on it. Unclaimed output is not automation. It is noise with extra steps. A place to fail visibly. If you cannot see the agent get something wrong, you will hear about it from a customer instead. A demo tests the component. Production tests the system. That is why the impressive pilot so often dies quietly a few months later. Usually nothing was wrong with the agent at all. The four things around it were never built. So the useful question is not which agent should I add next. It is which of those four is missing.