Running one agent is useful. Running several is where an agent harness earns its place on your machine, and it’s the reason Termic exists in one window instead of a pile of terminal tabs you lose track of.
Two kinds of parallel
There are two distinct shapes, and you’ll use both:
Many tasks, many branches. The classic fan-out. Open three worktree tasks, each on its own branch, each with its own agent, each working a different task. They can’t collide because each is a separate checkout on disk. This is the right shape when the tasks are independent: fix a bug here, add a feature there, bump a dependency in a third.
Many agents, one branch. Inside a single task, ⌘T opens another tab. You can run a second agent (even a different CLI) against the same working tree. Useful when you want one agent implementing while another reviews, or when you’re comparing how claude and codex each approach the same file.
Forking an attempt
When an agent is partway through something and you want to try a different direction without losing its progress, use Duplicate worktree from the task menu. It seeds a new task from the current branch tip, so you get a second agent starting from exactly where the first one is. Now you can let both run and keep whichever result you prefer.
Keeping track of them
The failure mode of parallel agents is losing track of which one needs you. Termic’s answer is the work-done indicator: a blue dot lands on a tab the moment its agent finishes a turn, and an orange bell appears when an agent is blocked waiting for your input. You glance at the sidebar, see what’s done and what’s stuck, and go to the one that needs attention. Optional desktop notifications tell you even when Termic isn’t focused.
The sidebar keyboard navigation is built for this: ⌥⌘↑ / ⌥⌘↓ hop between tasks, ⌘[ / ⌘] move through them in sidebar order, and ⌥↑ / ⌥↓ walk every visible row including each task’s tabs. See Keyboard shortcuts.
Ask everyone at once
When you want the same thing from multiple agents, Broadcast (⇧⌘B) sends one prompt to every agent in a task concurrently. It’s built for comparison: ask a question, let three agents answer, read the differences.
Running them safely
The thing that makes unattended parallelism actually relaxing is the sandbox. With it on, each agent runs in its own macOS Seatbelt cage with a network allowlist, and its per-action permission prompts are skipped, because the cage is the boundary. You can let several agents run with permissions off and know none of them can read your secrets or reach a host you didn’t allow.
When parallelism pays off
Be honest about this: parallel agents help when you have genuinely independent work, or when comparing approaches is worth the tokens. They don’t magically make one task finish faster. The day-to-day value of Termic holds up even at a single agent (tabs, editor, diff, run panel, resume); parallelism is the ceiling, not the floor.
Related
- Tasks & worktrees: the isolation that makes parallel safe.
- Work-done & notifications: knowing which agent needs you.
- Broadcast: one prompt, many agents.
- Scheduled messages & Ralph-style loops: keep one agent moving across several turns.