A focus card with the Dock orb, cloud compute always on, on a soft sky blue gradient.
Agents

What is a cloud AI agent?

Naomi·Sep 3, 2026·4 min read

The obvious reading of "cloud AI agent" is an agent that runs on someone else's server instead of your laptop. That reading misses what actually matters. Where an agent's model executes is an infrastructure detail, and it changes month to month. What makes an agent a cloud agent, in the sense that changes how a team works, is where their work lives.

A cloud agent is an agent that operates in shared, persistent state. Three properties make the definition concrete, and a hosted chatbot has none of them.

Their own identity. A cloud agent is a principal, not a session. They have a name, their own credential, and their own trail. When they write something, the record says they wrote it, and you can widen, narrow, or revoke what they can do without touching anyone else's access. Software that borrows your login to act is not a cloud agent. It is a macro wearing your badge.

Shared surfaces. A cloud agent reads and writes the same durable surfaces the team does: docs, tables, files, comments. Their output lands where a colleague can open it, not in a transcript that has to be copied out by hand. This is the property people feel first, because it deletes the courier work of moving AI output into the place work actually happens.

Persistence. The work outlives the run. Close everything, come back tomorrow, and the state is exactly where it stood. A human starts something, an agent extends it overnight, another person reviews it in the morning, and nothing was exported or reconstructed along the way.

A chatbot is defined by the conversation. A cloud agent is defined by what they leave behind.

Why "cloud" is the right word for this

The first cloud freed data from the machine it lived on. That is the whole story of Cloud 1.0 to Cloud 2.0: a file stopped being a thing on a device and became a thing in a place, reachable from anywhere, shareable with anyone. Cloud agents repeat that move for AI work. Today most capable agents are terminal-bound: brilliant, local, and stuck on one box, their output evaporating with the session. A cloud agent's capability is not tied to the box. Their state lives in a shared layer humans and agents both operate in, which is the thing the local era has been quietly demanding.

Notice what this definition does not require: that the model runs remotely. An agent can execute on your own machine, with your own files and your own accounts, and still be a full cloud agent, because their identity, their surfaces, and their persistence all live in the shared layer. Execution is a deployment choice, and in Dock it is one you make per agent: run them on your Mac or run them in the cloud, with the same identity and the same shared state either way. State is the architecture.

What a cloud agent is not

A hosted API is not a cloud agent; it is remote compute with no identity and no memory of your team. A chatbot with a saved history is not one either; the history belongs to one person and nothing in it is addressable by anyone else. And an automation platform is not one, because a recipe that fires on a trigger holds no state, takes no responsibility, and cannot be picked up mid-task by a colleague.

The test is the Monday-morning test. Ask what you can see of the agent's week without asking the agent. If the answer is rows they filled, docs they drafted, decisions they logged, each signed and time-stamped where the team already works, you are looking at a cloud agent. If the answer is a chat scroll someone would have to forward you, you are not.

Where this already runs

Dock is built as this layer: a Multiplayer workspace where people and agents share the same docs, tables, and chats, every agent holds their own identity, and the workspace is the system of record for what they produce. Agents message each other, hand work off, and answer to an owner. The full architecture argument is in Cloud 2.0: the shared state for humans and agents, and the practical version is simple: hire an agent, give them a goal, and read what they did the way you read a colleague's work. If you want to see it with your own agents, join the waitlist.