In Antigravity, connected to Colab T4 GPU runtime (confirmed by !nvidia-smi), AI agents like Claude Sonnet 4.6 cannot execute code on GPU. Manual runs in Antigravity notebook use GPU fine (e.g., 20 iterations fast), but agents fallback to CPU or fail with port rejections/Antigravity token issues.
Port checks show Antigravity Consumer Service needs auth; kernel uses internal WebSocket proxy with Antigravity tokens, but agents can't inject/access GPU.
Related to #388 (replied there), but specific to AI agent integration in Antigravity. Screenshots attached. What can be done to give agent the access to the gpu runtime?
(Might be extension of this issue - #388)