I’m absolutely exhausted with this setup. The state-management and model-switching logic in the new Antigravity IDE is actively breaking my projects, and the quota limitations make it nearly impossible to get any momentum.
Whenever the incredibly flimsy quota for Gemini 3.5 Flash runs out, I switch the agent over to Claude. Claude picks it up perfectly, no messing about, carries right on with the task, and honestly handles the execution better.
The nightmare happens when I return to Gemini. Instead of picking up where Claude left off, Gemini completely ignores every file change, terminal execution, and structural update Claude made. It acts like it’s stuck in a time capsule, reverting back to whatever state the workspace was in the exact second the quota hit. Because the agent doesn’t sync across the model switch, it ends up trying to rewrite or patch outdated code, actively breaking the project.
To make matters worse, Gemini is burning through its own limits just trying to figure out what’s going on. Just now, I watched it burn 50% of my quota because the agent got stuck in a doom loop. It analyzed the exact same file over 30 times, breaking the code, and then wasting tokens trying to fix the very errors it just introduced.
Google and the rest of the industry have spent months aggressively pushing the vibe coding paradigm onto non-coders and builders. We listened. We learned the workflows. We integrated it into our evening routines. I work a full day, and when I get home, I just want to sit down and build my own fun projects.
Having to constantly fight against aggressive, arbitrary quotas completely kills that workflow. If an agent stops halfway through a multi-file task because a tier limit dropped, you’re completely screwed. Because you’re working at a task-oriented level, you don’t instantly know what broke, what was fixed, or what state the terminal was left in.
I’m going to give it one last chance and try upgrading to Ultra, but I seriously doubt it will actually solve the core issue. Even if the tier increase gives a 5x multiplier, it isn’t enough if the agent is going to loop itself into oblivion. More importantly, will upgrading actually fix the context synchronization across model switches?
If switching models or hitting a limit means the IDE completely loses track of the project’s state history, then the agent-first architecture is entirely pointless.
Please sort this out. If you’re going to bait users with increased tier plans, the underlying state management needs to actually support a multi-model workflow without melting the codebase.