Most AI products begin with a chat box. That is useful for answering a question, drafting a document, or calling a tool. But it becomes harder to use when the work lasts longer than one conversation or involves several people and Agents. The problem is not only memory. The work also needs a clear goal, an owner, a current state, decision rules, review points, and a result that somebody can verify. That led us to a different model: an independent, goal-driven AI Workspace. What we mean by a Workspace A Workspace is not a folder containing several chats. It is an operating unit built around a measurable goal. For example, instead of asking an Agent to “help with product research,” a Workspace can define: the outcome it is expected to produce the people and Agents responsible for the work the Tasks and Plans currently in progress the Knowledge and documents that belong to this goal the Tools and Rules that apply the actions that require human approval the artifact that proves the work is complete Each Workspace keeps this context and execution state separate from unrelated work. A customer project should not inherit the documents, permissions, or unfinished Tasks of another customer. An internal research Workspace should not behave like a support Workspace. The goal provides direction; the boundary provides control. Why goal-driven matters Without a goal, an Agent can keep producing plausible activity without getting closer to a useful outcome. Inside a goal-driven Workspace, each proposed Task or action can be tested against a simple question: does this move the Workspace toward its goal? This creates a visible path: Goal → Plan → Tasks → Execution → Approval → Artifact The final unit of value is not a prompt or an answer. It is a completed outcome with enough history for a person to review what happened. Why independence matters Long-running AI work accumulates state. It gathers documents, decisions, intermediate outputs, failed attempts, permissions, and assumptions. If all of that lives in one global conversation, the system becomes difficult to understand and risky to operate. Independent Workspaces give each goal its own: Context Team and Agents Knowledge Tasks and Plans Rules and approval boundaries Activity history and artifacts This also makes collaboration clearer. A person can see what is active, what is blocked, what needs approval, and what has actually been delivered. What we built Manor AI is our self-hosted implementation of this idea. It connects Workspace conversations with Goals, Tasks, Plans, Agents, Skills, Knowledge, approvals, Flows, Automations, and resulting artifacts. The intention is to keep context, execution, review, and results together instead of scattering them across chat windows and separate tools. A small first workflow looks like this: Create a Workspace with a clear outcome. Add the relevant Knowledge and an Agent with a defined responsibility. Create a Task with acceptance criteria. Review the Plan and respond to approval requests. Verify the resulting artifact. Turn a process that works repeatedly into a Flow or Automation. We have published the self-hosted source on GitHub under the Manor Sustainable Use License 1.0. It is source-available rather than OSI-approved open source, and we want to be explicit about that distinction. Repository: https://github.com/manor-os/manor-ai What we want to learn We are still testing where the Workspace boundary should sit and what information must remain visible to users. If you build or operate Agents, we would appreciate your view on three questions: What should be isolated between Workspaces? Which actions should always require human approval? What evidence do you need before trusting an Agent's “completed” result? Those answers are more useful to us than a generic launch reaction.