Isolated sandboxes

Every agent in Agent Swarm runs inside its own cloud sandbox, branched from main. There is no local execution — all file operations, commands, and git happen inside sandboxes.

Why isolation matters

Isolation is what makes parallelism safe:

• A planner can read the codebase without a coder mutating it underneath.
• A coder can write and run tests without corrupting another agent's assumptions.
• Compute-heavy work — Playwright, builds, full test suites — doesn't bottleneck on your laptop.

main
 ├─ sandbox A  (planner)   read-only exploration
 ├─ sandbox B  (coder)     writes code, runs tests
 └─ sandbox C  (tester)    records video proof

What's in a sandbox

Each sandbox is a full Linux environment with:

• A clone of your repo at the agent's branch
• Installed dependencies and a running dev server
• A terminal you can stream and type into
• A desktop with an in-VM browser, so logins, OAuth, and cookies all work normally

The stale-state problem becomes a notification

When agent B merges to main, agent A receives a signal: "main moved, here's the diff since your branch point." The agent decides whether to reconcile (rebase, re-plan) or ignore the change if it doesn't overlap. This is cheaper than re-reading everything and removes the endless "go look at the code again" loop.

Keeping sandboxes warm

Sandboxes pause when idle to save resources and resume on demand. While a desktop, terminal, or app-log view is on screen, a keepalive heartbeat ensures the environment is never paused out from under you.