Three days ago I posted about the 503 / agent terminated errors disappearing for the first time in months. Figured it might be a fluke. It wasn’t.
I’ve been running heavy sessions daily since then. The errors are basically gone. Not completely zero, but close. Maybe one or two terminations across multiple full days of work versus the old rate of hitting it every few minutes.
The bigger change I noticed is the wait time. When the model does hit capacity now, Antigravity holds the connection open way longer before giving up. It used to terminate within seconds. Now it actually waits for a seat to free up on the server. Most of the time it gets through. That’s a real backend change, not just more capacity.
Between the reduced error rate and the longer retry window, the product is actually usable for sustained work again. Whatever shipped on the backend this week, it worked.