I am reporting a significant regression in the usability of the Google AI Pro subscription within the Antigravity IDE over the last three weeks. As a systems engineer, I’ve identified several architectural “token leaks” that make the current $20/month tier non-deterministic and functionally unusable for professional workflows.
Technical Observations & Pain Points:
-
Redundant Output Loops: The AG agent currently duplicates “Thinking” and “Planning” logs across both the main chat window and the dedicated agent window. This results in massive, unnecessary token consumption for meta-commentary before any code is written.
-
Verbosity Overhead: Forced “Technical Research & Planning” artifacts and long-winded walkthroughs consume the context window rapidly. These should be opt-in, as they currently prioritize “digital pacing” over execution.
-
The “Agent Exhausted” Reset Loop: Frequent
Agent execution terminatederrors (often after only 2–3 prompts) force the agent to lose its volatile state. Upon restart, the agent re-scans the entire workspace, burning the remaining quota just to regain context of the previous failed task. -
Inconsistent Quota Windows: We are seeing a shift from the standard 5-hour refresh to “dynamic” lockouts lasting 1.5 hours to several days. This makes it impossible to plan a standard development sprint.
-
“Google Only” Handshake Errors: Persistent authentication errors in the Antigravity IDE are interrupting active agentic loops, leading to wasted compute and incomplete file writes.
Requested Resolution:
-
Streamline Agent Logic: Implement a “Silent Execution” mode that bypasses verbose planning artifacts.
-
State Persistence: Ensure agents can resume from the last successful checkpoint after a “Google Only” or “Exhausted” error without re-calculating the entire workspace context.
-
Quota Transparency: Provide a clear, real-time indicator of remaining “Thinking” tokens versus “Generation” tokens so users can manage their sessions defensively.