Known gaps in the beta

[alchaplinsky]

Rather than let you discover these one at a time, here is what I know is missing or rough right now. If something here is a blocker for you, say so in a reply. That is genuinely how I decide what to fix next.

1. macOS only. The default isolation engine is macOS sandbox-exec, which has no direct equivalent elsewhere, so Windows and Linux mean building a different isolation story rather than porting a UI. On the list, not close.

2. Agents can publish under your identity. Your credentials never enter the sandbox: pushes and PR creation run on the host through a brokered channel, so an agent cannot read your token. But those operations currently run without a confirmation prompt. Choose tasks and repos accordingly. Adding a confirmation step is a known gap.

3. The default sandbox is write protection, not a VM. Under Seatbelt an agent cannot write outside its workspace, but it runs as your user, so it reads what you can read and the network is open. The opt-in Docker engine closes off reads too, and the network is still open there. Every harness except Antigravity supports Docker.

4. Usage analytics are on by default. Anonymous, random per-install ID, five events (app opened, app updated, agent spawned, turn completed, PR opened), and never your code, paths, repo names or prompts. One toggle in Settings, General. Tell me if you think the default is wrong.

5. Parallel work still integrates serially. Isolation stops agents colliding while they work, but it defers the collision to merge time. Five agents on five branches of one repo will conflict, and today you resolve that the normal way. There is no rebase orchestration, staleness indicator or merge sequencing yet. This is the unglamorous half of parallelism and I know it.

6. Review does not scale the way spawning does. Running ten agents is easy; reviewing ten diffs is ten reviews. What is missing is one review queue across every agent, where each change arrives with the evidence already attached: tests passed, reviewer findings, budget consumed, CI status. The pieces exist and do not yet feed each other.

7. Closing the app stops your agents. Workflows are resumable and carry wall-clock budgets, so they are built to outlive a session, but they still live inside a foreground macOS window. There is no headless mode yet.

8. Pi and Antigravity are experimental. They work, with more rough edges than the others.

Nothing here is a promise about dates. It is what I would want to know before pointing a tool at my own repo.