The agent can now work directly on your Windows machine through the approval-gated bridge. Every file write lands in a keep/revert review set: accept or revert per hunk, and your manual saves sync into the same queue.

Dangerous commands never auto-run, auto-allow lists stay local to your device, and every remote call is rate-limited and fully logged, and the CLI reports exactly what ran and why.

Complex tasks now fan out across up to 40 subagents running in parallel, all following the same model picker as your chat. Delegation, tool calls, and todo lists render live in the conversation so you can follow along.

On heavy workload, subagents may not be completely available; the main agent keeps going and picks them up again as capacity frees up. Better and more accurate results on complex work is the goal; raw speed is not guaranteed.

Subagents share the workspace's 50 GB upload budget, so large research sets and datasets move over once and stay available to every agent in the session.

A single prompt can now keep working for up to 24 hours. Your cloud workspace (2 CPU / 4 GB) keeps running with your PC off, and the companion mobile app lets you monitor progress and send follow-up prompts from anywhere.

Sessions, terminal output, and file changes stream back as they happen, so a long run never turns into a black box. If something needs your call, the agent asks and waits.

The cloud sandbox is powerful but not permanent: environments reset between sessions. Connect GitHub access and agents submit completed work to your repository as commits, branches, and pull requests, even while you're offline.

Work saved on the chat always persists, but the files live in the environment. GitHub is the bridge between the two: the durable home for everything your agents finish.

Pair it with the local bridge and you get the full loop: cloud work lands in your repo, local work lands on your disk, and you review both in one place.