At first, the agent was helping me build a new app and doing the verification of a plan. Because it is an Android app, the agent needs to run ./gradlew assembleDebug. However, it always gets stuck on "Working..." forever, without outputing any text.
I tried advices from Google search and Gemini, but nothing works.
After a while, I realized that the agent would get stuck on "Working..." even by running a simple command echo hello > test.txt. So I reframed my problem as a terminal command thing, not an Android/Gradle issue.
After another round of try-and-errors, I realized the problem is the Agent security mode.
By clicking the bottom right button “Antigravity → Settings” → “Advanced Settings”, there are three modes: Full access, Sandboxed, and Strict.
I figured if I use Full access, the agent will be able to run echo hello > test.txt. But when I switch to Sandboxed mode, the agent will be stuck at "Working..." forever without outputting any text.
I’ve tried the following, but nothing works:
Add "terminal.integrated.windowsEnableConpty": false to the settings.json
Switch default terminal profile to PowerShell/Command Prompt.
Add cmd /c at the begining of the command
Add < NUL at the end of the command
Add global rules telling the agent do not do interactive stuff, or make sure to end properly, etc.
Reinstall Antigravity IDE + Antigravity
Remove %USERPROFILE%\.gemini\antigravity\ and %APPDATA%\Antigravity IDE\
I really want to use Sandboxed instead of Full access, so that is not actually a fix.
Below is my environment:
Antigravity IDE Version: 2.1.1
VSCode OSS Version: 1.107.0 (user setup)
Commit: e0b7a2bcf575cfba10528c4e7c10bd3ce2d7769a
Date: 2026-06-17T22:39:43.962Z
Electron: 39.2.3
Chromium: 142.0.7444.175
Node.js: 22.21.1
V8: 14.2.231.21-electron.0
OS: Windows_NT x64 10.0.19045
Language Server CL: 933924006
considering after a few turns it stops even asking for approvals and just makes changes, eventually even ignoring the actual sandbox and then the current sad state of functional interconnectivity of Antigravity and the IDE, as well as the most recent show of token burning with failed and looping thinking steps and worsening contextual management… the only thing google seems to be investing in with regards to antigravity is in secretly eating quota and then pushing more and more to upgrade plans or enable overage.
Like you I’ve observed over the last few days that it’s not able to even successfully run sandbox-friendly commands without first doing a bunch of thinking then showing an error, then thinkibg some more trying some more things ultimately just hanging or producing a regressed code (and the accept/reject diff review doesn’t mean anything, it’s already saved the new changes and reject/reject all doesn’t revert.)
And it’s unclear if this behavior you’ve observed which i definitely also have over the past few days (as well as running out of quota within 10 turns that only actually ruin my working/uncommitted codebase regression after regression while looping the same thinking steps so fast that you can’t even stop it in time to avoid quota impact) is just buggy behavior or a pattern of consumption and demand for more pay to continue… you’re not alone, and it doesn’t seem to be a concern for google.
you sound so much more frustrated than I feel you to some degree
yeah I noticed that the UX in Antigravity regarding that Agent security mode is very strange.
First, if you freshely install the Antigravity IDE, and you open up the settings, you would see the default mode is Full access. WHYYYY??? That does not sound very safe.
Second, I notice sometimes even if I set the option in “Terminal Command Auto Execution” to “Proceed In Sandbox”, and then I close the settings, and reopen it, the option would be set to “Always Proceed” again.
The token burning situation does not apply in this case. In this case, it just gets stuck there, no more tokens burnt. But I think I’ve run into a situation where the agent cannot connect to the internet to call the API properly, so it just keeps using up tokens until my quota ran out.
I feel like it’s not that Google is pushing us to burn more tokens or something, it’s more like, probably the code in Antigravity IDE (or I should say, everything) is written by AI, and they don’t put enough effort to make sure each little feature actually works as expected, because they are too busy using AI to be “more productive”.
That is all I get as output after disabling sandbox. I started having huge delays with Antigravity taking over 15 mins to a simple “hi” message. Stuck in “working” forever and then producing awful results. I tried deactivating the sandbox and now it is still taking a long of time but outputs just a bunch of “producing” concatenated together, which I am pretty sure might be burning my tokens while doing nothing. AI companies are turning into an scam with their business model.
I tried reporting this on my setup using Sandboxed mode, including creating files, creating folders, and running simple terminal commands (such as echo hello>test.txt), but they completed successfully and i wasn’t able to reproduce the agent getting stuck on “Working…”
Could you please provide a bit more information?
Which AI model were you using when this issue occurred (e.g., GPT-OSS 120B, Gemini 3.6 (High), Claude, etc.,)
Does it reproduce in a new empty workspace, or only in your Android project?
What is the exact prompt that consistently triggers the issue?
Could you share a screen recording of the issue and, if possible, the IDE logs/Developer Tools console?
I just created a new folder (folder name “new”) and found the problem is still there.
AI model used: Gemini 3.6 Flash (Medium)
I also tried Gemini 3.5 Flash (Medium). Same issue.
Exact prompt: run: echo hello > test.txt (as you can see from the first screenshot)
I also tried similar prompts like: try running echo hello > test.txt, or run: and then the next line echo hello > test.txt
but the results are all the same.
While the screen gets stuck at “Working”, the Developer Tools repeatedly shows two errors. You can see more details in the second screenshot. (I tried to Google/asked AI about these errors but couldn’t find the fix either.)
