Caspian 東澔
2.5 KiB
Parallel agents
Helmor's whole premise: agents are fast, you are the bottleneck. The fix is not a faster agent — it's running several and getting good at switching.
The rhythm
- Fan out. Create a workspace per task (⌘N) and dispatch each one. Isolation makes this safe — five agents in five worktrees can't conflict (Workspaces).
- Rotate. While agents work, you review. Move to whichever workspace needs you, read the diff, steer or approve, move on.
- Ship and archive. Merged work leaves the sidebar; new tasks enter.
Knowing where you're needed
The sidebar is your control tower:
- Running — the agent is mid-turn; nothing needed from you.
- Needs input — the agent asked a question or wants permission.
- Unread — finished while you were elsewhere.
- A glowing border means run scripts (dev server, tests) are active.
Desktop notifications (with optional sounds) fire when an agent finishes or needs you — configurable in Settings → General.
Switching fast
- ⌃Tab — quick-switch overlay between recent workspaces (hold ⇧ to go backwards)
- ⌘⌥↑ / ⌘⌥↓ — previous / next workspace in the sidebar
- ⌘⇧F — filter and sort the sidebar
- ⌘L — jump focus straight to the composer
Organizing the fleet
- Group the sidebar by repository or by status, or keep a flat list.
- Status lanes — assign workspaces In Progress, Review, Done, Backlog, or Canceled and group by status for a kanban-style view.
- Pin the workspaces you keep coming back to; mark unread anything you want to revisit; drag to reorder.
Drive-by tasks
The quick panel (⇧⌥Space) is a small floating window for tasks that don't deserve a context switch — fire off a prompt from anywhere, glance at the response, and either dismiss it or open it fully in the main window.
Dispatching from the terminal
Parallelism composes with your shell. The CLI can create workspaces and send prompts without touching the GUI:
helmor workspace new --repo myapp
helmor send --workspace myapp/refactor-auth "Extract the session logic into its own module."
See CLI & MCP.
How many is too many?
In practice the limit is your review bandwidth, not the machine. A good starting shape: two or three substantive tasks plus a couple of small ones. Plan mode helps you keep risky work cheap to supervise — let the agent propose before it implements.