- InfoWorld's hands-on review describes Google's new Antigravity 2 app as agent chat with no code editor, built around projects that each have their own permissions. Google still offers the Antigravity IDE.
- WebPulse argues that without an editor, oversight may shift to rules set before the agent starts. The review did not test whether those rules hold.
- Ask which projects agents may modify, which mode they run in, and which models your policies allow.
When a tool removes the code editor, it removes the editor through which a person would normally watch the work. Google's new Antigravity 2 app does exactly that. The lesson here, in WebPulse's reading, is that supervising an AI agent may stop being about reading code as it appears. It may become about the rules you set before the agent starts. The facts below come from InfoWorld's review.
What Google changed
InfoWorld's Serdar Yegulalp tested Antigravity 2 hands-on. Google describes it as a "standalone desktop command center" for agents. Yegulalp writes that it is closer to LM Studio, a desktop app for working with models, than to a full integrated development environment (IDE), the software where programmers write and debug code.
The app is chat-centric and has no code editor or debugging tools. Google has not retired the IDE version. Teams that want editing and agents together can stay with it. The new app is aimed at people who prefer chat alone, or who want to keep hands-on development separate from what the review calls vibe coding.
How the controls work
The main change is how work is fenced in. Earlier versions of Antigravity used workspaces. Those were confined to one directory and lacked granular permissions or customizations.
Antigravity 2.0 replaces workspaces with projects. A project can span several directories that the user defines explicitly. Each project has its own isolated settings and permissions. Yegulalp says this suits software that has shared components in different places.
A second choice is between local mode and worktree mode. In local mode, the agent edits your actual folders or Git checkouts in place. The review links that to convenience and speed. Worktree mode creates a new Git worktree, a separate working copy of the code, for each conversation. The active working folder is not affected, and subagents running in parallel do not conflict.
Subagents are helpers spawned by the main agent. Each works on one task in its own isolated workspace. Agents or subagents can also run in the background, so the user and the main agent keep working. Yegulalp notes growing interest in using parallelism to speed up agentic development. He says Antigravity offers two complementary routes to that: subagents and background runs.
Why the permissions screen may matter more
This is WebPulse's interpretation, not a finding of the review. In an IDE, a person can see each file change. In a chat app, the person sees the conversation. The review does not say whether changes can be checked by other means. If they cannot, the settings that say which directories an agent may touch, and whether it works on a copy, become the main safeguard.
The review's own example hints at the stakes. One screenshot shows a chat about a Python web project that the agent has permission to modify. Yegulalp did not test whether the permission boundaries hold under pressure. This is one reviewer's hands-on account, not a security assessment.
Where the review found limits
Yegulalp reports that Antigravity's API timed out during a /grill-me session. That command has the agent ask the user focused questions about the project. After the timeout, the context vanished and could not be recovered. He adds that the tool tends to forget context, including questions already asked, after such a timeout.
He also names what he calls the biggest limitation. The app works only with versions of Gemini, Claude and GPT. It has no way to connect to a self-hosted model through an industry-standard interface. He bets this is by design and points readers who need that to LM Studio.
Questions to put to your team
If developers use agent apps like this one, five questions are worth asking. Which projects may an agent modify, and who sets those permissions? Do agents on shared code run in worktree mode or local mode? Who reviews what subagents produce before it merges? Do the three supported model families fit your policies on where prompts and code may go? And what is the recovery plan when a session fails mid-task and the context is lost?
An agent without an editor still needs a supervisor. In Antigravity 2, that job may well begin at the permissions screen.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.
