Dock starts with a problem that sits underneath chat: work needs owners, shared artifacts, memory and a handoff, not only another answer.
The easiest way to describe a new AI product is to start with the model. Which model does it use? How fast is it? How much context can it hold? Those questions matter, but they are not the place to start. I start with a more ordinary problem: work gets split across conversations, files, calendars, messages and people. A prompt can produce a useful answer inside one conversation. It does not, by itself, give that answer an owner, a place to live, a next step or a way to reach the person who needs to review it. That gap is what I was trying to change. Not the existence of chat. The position chat occupies in the work. A prompt is not a unit of work Dock is a workspace where proactive agents work alongside people. An agent has a name, a role, objectives and a reporting line. It can work in the shared workspace, use the access granted to it, keep a memory of the work and hand a piece of a goal to another agent. Those details sound like configuration until you compare them with the alternative. In a chat window, the person usually remains responsible for carrying the answer into the next system. The person copies the context, remembers the decision, follows up later and checks whether anything happened. The model may have done the fast part, while the human has become the workflow. I do not think adding more chat windows solves that. It can make the middleman faster, but it does not remove the middleman. The team is the product boundary Mike Molinet describes that experience directly in the launch post: he had opened five chat windows and was carrying context between them. He wanted something closer to a team than a smarter chat window. The useful product question is what a team requires. It requires ownership. If an objective belongs to an agent, somebody should be able to ask what changed, what is blocked and what decision is needed. It requires a shared surface. A document, table, file or thread should not disappear into a private transcript. It requires coordination. Work should move through a handoff rather than through a person’s clipboard. Dock positions a company of agents around these structures, and the phrase is only useful when it points to them. Names and roles are not theater. They are a way to make responsibility visible. For example, imagine a content agent preparing a launch draft and a chief-of-staff agent collecting the approvals around it. The useful boundary is not that two personalities are pretending to be colleagues. It is that the draft, the open questions and the approval request are visible in the same workspace, with a defined owner for each next step. Proactive does not mean unsupervised Dock’s public description includes schedules, daily logs, guardrails and escalations. An agent can work when a person is away and return with what changed. That is different from saying that every action should happen without review. The boundary is explicit: an agent should ask before anything irreversible or outward-facing, escalate when a decision belongs to a human and write down what it did so another person can pick up the work. Proactivity is about starting the next useful step. It is not permission to make every consequential decision. That distinction matters for the product. The aim is not to make people disappear from work. It is to let people spend less time transporting context and more time deciding, correcting and directing. The system should outlast a model choice Dock describes support for Claude, OpenAI, Kimi and GLM, with a model selected per agent and a cost setting. The point is not a permanent ranking of those models. The point is that the surrounding system—memory, role, access, rules and workspace—should not have to be rebuilt every time a model improves or a team chooses a different one. That is a practical design constraint. A model is one layer. The work also needs a runner, computer and storage for an agent, credentials in a vault, channels such as email and calendar, and native documents, sheets and files. Without those layers, a strong answer still has nowhere to go. Dock is not a claim that agents are people or that work has become automatic. My earlier essays on Agent Washing and The Era of Agents and Humans explore the broader shift. This series is narrower: what the product has to hold together if that shift is going to be useful. The change I was trying to make is simple to state. A person should be able to give a goal to an accountable agent, see the work in the place the team already uses, receive an escalation when judgment is required and return later without reconstructing the entire story. That is a different center of gravity from prompt-by-prompt work. It is also a more demanding standard than a good answer. Source note: This essay draws on Mike Molinet’s Dock launch article and my Dock agent explainer, both published September 9, 2026.