Workspace: conversations as a tree

Workspace is a branching agent workspace where user prompts and agent turns are nodes in a tree, rather than a single linear transcript. You can fork from any turn and that branch's ancestors and explicit references in the prompt are sent to the model. Branching gives control over context that you can't get with a linear chat, and makes it easier to research and explore different solutions without fear of poisoning the context.

These screenshots follow a small task app called Harbor. The first prompt asks the agent to search the repo and explain how it stores tasks. Click any image to see it full size.

Workspace's empty canvas with a prompt asking the agent to search Harbor's repo and list tasks with their status. — 01Starting with a prompt.

Context and branches

Workspace maintains embeddings for documents, code and chats in the project so agents can find information outside their branch and build on common goals. Here the agent searches the repo and reads the task code. Opening the card shows the tool calls and response.

Harbor's first turn on the canvas, with repo search results and a read of src/tasks.py open in the inspector. — 02Searching the repo and reading the source.

Forking starts a child turn with the context up to that point. The next branch adds a function to count open tasks and writes notes about what to build next. Agents can also spawn subagents with the same branching context logic.

Working on files

Each project has its own working directory, shared across branches. Agents can search, write and edit documents and code, execute code and run commands, and read/write to GitHub. Branching separates the conversation, not the files. An edit on one branch is still an edit to the shared project.

A child turn shows changes adding count_open to the task code and creating notes/build.md. — 03Editing the code and writing build notes on a child branch.

The inspector shows a Python command calling count_open and returning 2, followed by a summary of the changes. — 04Running the function. Two tasks are open.

A separate branch can go off and research something else. Here it searches the web for the latest FastAPI release. That conversation doesn't become part of the code-editing branch's history unless you explicitly bring it in.

Sibling branches for task code edits and FastAPI research, with web sources open in the inspector. — 05Web research on a sibling branch.

Charts and documents

The outputs are files in the project too. An agent can run code to produce a chart or write an HTML page, and you can open either beside the conversation. Here it charts the task statuses and makes an HTML page for Harbor.

The image inspector displays charts/tasks.png, a bar chart of Harbor's open and done tasks, beside the conversation tree. — 06The generated chart, opened from the project files.

The HTML inspector renders docs/harbor.html with Harbor's task board, beside its design card on the canvas. — 07An HTML document rendered in the inspector.

The project is split into a UI library, an open-source local app you can point at a directory like VS Code or Cursor, and a closed-source app that hosts cloud agents for long-running tasks you can access from multiple devices.