Is Antigravity (AG) IDE an example of poor QA in AI-assisted coding tools?

I’ve been using Google’s Antigravity IDE since before v1.16.5, and I’ve noticed a significant decline in stability after that version. It feels like releases are being pushed without adequate smoke testing or regression checks.

Issues introduced after v1.16.5:

  1. Chat history loss — This is the most critical bug. Entire conversation histories disappear. How does something this fundamental pass QA?

  2. “Always Run” regression — The “Run Command > Always Run” setting stops working, even when Turbo Mode or “Always Proceed” is enabled. This was previously fixed and has regressed.

  3. Browser task regression — Same issue as above — “Always Proceed” is ignored for browser-based tasks.

  4. Extension crashes — Multiple extensions fail or crash on the latest version. While not entirely AG’s fault, it suggests breaking changes were introduced without sufficient compatibility testing.

  5. Workflow folder migration with no notice — The workflow directory was silently changed from global_workflows to .agent/workflows with no migration path or user notification. This is a terrible upgrade experience.

Forced update to v1.19.6:

I stayed on v1.16.5 specifically to avoid these issues, but today AG forced an update to v1.19.6. Now I’m back to manually clicking “Run” and “Allow” for every command and browser action — even as an Ultra subscriber with all auto-proceed settings enabled.

I appreciate the product and want it to succeed, but the current release process feels rushed. Basic regression testing would catch most of these issues. As a paying Ultra user, I expect better quality control.

Has anyone else experienced similar problems after updating past v1.16.5?

I am writing to express my severe disappointment with the quality of this service. Given the resources behind this company, the current user experience is unacceptable. Despite being a paying customer, I am dealing with constant bugs, slow speeds, and quotas that aren’t being properly allocated. I integrate this heavily into my development work, but the constant troubleshooting makes it feel like I’m dealing with a broken piece of equipment rather than a helpful tool. This needs to be addressed

1 Like