Agent Mode and Editor Mode¶

Cortex Code Desktop runs in two top-level workspace modes — Agent Mode for running and reviewing many agents across workspaces in one window, and Editor Mode for the full VS Code-style IDE scoped to a single workspace.

Modes at a glance¶

Agent Mode¶

Agent Mode replaces the editor area with an agent-first surface. The conventional editor parts (activity bar, primary side bar, editor groups, panel) are hidden so the chat and its surrounding context get the full window. Agent Mode is designed around three capabilities:

1. Parallelism across workspaces, in one window¶

Editor Mode is one workspace per window. Agent Mode lets you drive agents in many workspaces side-by-side without juggling windows.

• The sidebar lists all workspaces you’ve added, each with its own conversation history.
• Switching workspaces is a single click — no Open Folder, no window reload.
• Each conversation runs in its own workspace context: file pickers, terminals, search, and tool calls all scope to that conversation’s folder.
• A built-in Playground workspace is always available for one-off chats that don’t belong to any project.

2. Built-in review of agent changes¶

Each conversation tracks the files an agent has modified. Agent Mode surfaces these changes directly, so reviewing and acting on them does not require opening separate source-control or diff views.

• A Files changed panel in the side rail lists every file the active conversation has touched, with inline diffs and per-file accept and reject actions.
• Pending changes, approval prompts, and diff stats are scoped to each conversation and persist as you switch between them.
• Bulk actions — accept all, reject all, open file — are available without leaving the Agent Mode surface.

3. Background execution and notifications¶

Conversations continue running when not in focus. Agent Mode’s sidebar acts as a status dashboard for everything in flight at once — Editor Mode’s chat panel shows only the active session.

• The sidebar shows per-conversation state for every workspace at once: in progress, awaiting approval, unread.
• Approval prompts and notifications attach to the conversation that needs attention, so you can triage many in-flight agents without switching focus.
• Scheduled tasks appear in the same list, with the same indicators.

Surfaces at a glance¶

• Sidebar — your workspaces and their conversations, with search and a button to start a new agent.
• Conversation pane — the active chat, including tool calls, approvals, and inline diffs.
• Right-hand rail — Files changed, terminal, and other inspectors for the active conversation.

Editor Mode¶

Editor Mode is the familiar IDE: file explorer, editors, source control, run & debug, terminal, and extensions are all front-and-center. Chat lives in a side view rather than as the main pane, so you can edit code and talk to the agent simultaneously.

• All editor-area features (multi-cursor editing, find/replace across files, language tooling, debugging, tasks) are available exactly as in upstream VS Code.
• Extensions, keybindings, and settings carry over — Editor Mode is not a stripped-down view; it is the full workbench.
• The chat side view shares the same conversation list as Agent Mode, so you can pick up where you left off.

Workspace scoping¶

Conversations, skills, and plugins are all scoped per workspace. Agent Mode and Editor Mode are different views over the same underlying state, not separate stores.

• Conversations belong to the workspace they were started in and persist in that workspace’s storage. Editor Mode shows the current workspace’s conversations; Agent Mode shows conversations across every workspace you’ve added, grouped by workspace.
• Skills and plugins resolve from a global set (shared across all workspaces) plus the active conversation’s workspace-local set. An agent running in one workspace only sees that workspace’s project-defined skills and plugins, never another workspace’s.
• Tool calls (file picker, terminal, search) always run in the active conversation’s workspace context, regardless of which mode you’re viewing it from.
• Starting, continuing, archiving, or deleting a conversation in one mode is reflected in the other.

When you toggle, your active conversation follows you only if it belongs to the workspace Editor Mode is open on. Conversations from other workspaces remain accessible in Agent Mode’s sidebar.

Choosing a mode¶

Rule of thumb: if you are writing the code, use Editor Mode. If agents are writing the code and your job is to dispatch them, watch them, and review their output — especially across more than one workspace — use Agent Mode.

• Hand-editing, debugging, using extensions — Editor Mode. Keep the agent in the loop via the chat side view.
• Driving one or more agents on real work — Agent Mode. Chat gets the main pane and the side rail surfaces the agent’s changes, terminal, and other context as the work progresses.
• Running agents across multiple repos / workspaces in parallel — Agent Mode. One window covers all of them.
• Long-running or scheduled tasks — Agent Mode. Unread indicators and notifications are designed for stepping away and coming back.