In my O9X BuildLog 11, I established a manual “Tri-Model” workflow: Gemini as Architect, ChatGPT as Auditor, Claude as Builder. The logic was sound. Checks and balances between models keep quality up. The execution was painful.
Every time I moved context between browser windows, I risked losing data. I was acting as the copy-paste bridge between three separate systems. That is not a workflow. That is a tax.
The Shift
I have moved the O9X build environment from VS Code to Google Antigravity. The three-window manual loop is now a single IDE with the same model diversity baked in.
I do not switch tools for features. I switch for velocity. The Tri-Model workflow was right in principle but slow in practice. Antigravity removes the integration friction without removing the model variety.
Here is how the new setup works. I run two asynchronous agents in the same pane. Agent 1 is Gemini 3 Pro, holding the Architect context — the schema and the product requirements document. It has read access to the repo; I restrict its write access. Agent 2 is Claude, the Builder. I give it specific execution tasks. For non-critical logic checks, the IDE’s integrated access to GPT-OSS-120B handles a fast, cheap second opinion without burning paid inference credits on the main models. If the check is critical, I open a fresh ChatGPT window.
The Browser Agent Trap
Antigravity ships with a Browser Agent that can theoretically drive the UI to test features. In practice, it fails at the gate. My features sit behind a login and MFA wall. The agent cannot navigate session tokens consistently.
My response was simple: I do not fight the tool. I test manually. While the AI writes the code, I am the quality assurance. I click the buttons. I see the latency. Keeping the testing loop manual means I retain direct control over the user experience. That is not a concession. It is a deliberate choice.
What Actually Changed
The old workflow was one IDE, two browser windows, and manual copy-paste. The new workflow is one IDE, one browser window when needed, and split panes with me in the loop for testing. The supervision that used to happen across windows now happens in real time. I am no longer the bridge between systems. I am the referee inside one system.
The switching tax is gone. The model diversity stays.