Most teams adopt AI agents one session at a time. Each person opens a chat, pastes some context, and drives. That works for a single task, but it leaves the operation itself nowhere: the process lives in prompts, the decisions live in chat history, and the record of who approved what lives in nobody's system. An operations layer is the missing tier that turns many separate sessions into one run operation.
It is different from the layers around it. The model layer (Claude, GPT and others) produces text and tool calls. The client layer (Claude Code, Cursor, ChatGPT, Codex) is where a person or a runner talks to the model. Agent frameworks are code you write to wire models to tools. An operations layer does none of that. It holds what those layers forget: the defined process, the state of every piece of work, the checkpoints, the shared knowledge, and the audit.
A useful way to test whether you have one: could a new agent, or a new colleague, pick up a half-finished piece of work tomorrow and continue it at the right step, under the same rules, knowing what the last run learned? If that depends on someone remembering, the operation is still in people's heads.