What they did
Gwerzman uses Claude Code as the lead agent, with Fable for design, Opus for code and Sonnet for tests, and Google's Antigravity running Gemini 4 Argon as the reviewer at every stage. For a long time he carried reviews between the two windows by hand. In November 2026 he and Claude Code built agy-bridge, an open-source MCP server that drives Antigravity's headless agy CLI. Claude briefs a reviewer agent, the reviewer reads the project, runs the tests and returns a structured verdict with findings and questions, Claude fixes or argues each point once, and the fix goes back for a re-check. The bridge backs up uncommitted work before each reviewer turn and reports any file the reviewer changed. Gwerzman is called in only when the two agents still disagree.
How it went
Reviewing the bridge's own design and code, Antigravity raised two high findings Claude had passed: two terminals could start competing review processes, and nothing saved uncommitted work before the reviewer ran shell commands. Both were real and cheap to fix. In another round the reviewer found a real restore bug but proposed a fix that would have skipped patches from subfolders; Claude tested it in a scratch repo, pushed back and the reviewer agreed. In one test run Claude skipped the re-check of its own fixes, so the skill now forbids that.
Worth knowing
He insists on two different tools and two different models, since Antigravity can also run Claude and that would share Claude's blind spots. The reviewer gets a shell, so a stray command could still touch files; the backup and compare step is the safety net. A reviewer also follows skills found in the repo's .agents folder, so look before reviewing code you do not trust. Each review turn takes minutes, so he skips it for throwaway work.