Why every agent gets its own sandbox
Isolation isn't a nice-to-have for parallel agents — it's the thing that makes them safe. Here's how we think about it.
If you've ever run two agents against the same working directory, you know the failure mode: one edits a file, the other reads a half-written version, and both end up confused.
Isolation makes parallelism safe
In Agent Swarm, every agent runs in its own cloud sandbox, branched from main. That single decision unlocks a lot:
- A planner reads the codebase without a coder mutating it underneath.
- A coder runs the full test suite without slowing anyone else down.
- Heavy work — builds, Playwright, large installs — never touches your laptop.
main
├─ sandbox A planner
├─ sandbox B coder
└─ sandbox C tester
Stale state becomes a notification
The hard part of parallel work is staying in sync. When one agent merges to main, the others get a signal: main moved, here's the diff since your branch point. Each agent decides whether to reconcile or ignore it.
That's much cheaper than re-reading the whole repo, and it removes the endless "go look at the code again" loop that plagues single-context agents.
The feedback loop stays tight
Isolation usually means things feel remote. We fixed that with a streamed desktop, a live terminal, and branch-aware diffs — so even though the work happens in the cloud, it feels like it's right in front of you.
Read more in the Isolated sandboxes docs.