Google has open-sourced an Antigravity plugin that makes AI coding work pass through a written design and review process before implementation. Its central trick is to give fresh AI reviewers the plan without the conversation that produced it, a neat attempt to keep the original team’s blind spots from travelling with the code.
Watch Desk analysis
What happened
The Elephant-Goldfish Model plugin, or EGM, separates planning from implementation. A long-running “Elephant” session discusses the problem with the developer and keeps the project’s architecture and decisions in view, but writes no code. The team then drafts a Markdown design document covering the problem, technical plan, alternatives and specific files to change.
In the review phase, stateless “Goldfish” sessions check the document. One must explain the design back to the Elephant; if the explanation exposes ambiguity, the plan goes back for revision. Other reviewers look for edge cases, hidden dependencies and logical gaps, and a final reviewer checks whether the file paths, data structures and function signatures are clear enough to proceed. Help Net Security’s report says Google has released the plugin free on GitHub, that it works in Antigravity, and that support for other environments is coming.
Why it matters
This is a concrete answer to a familiar problem with AI coding tools: they can make producing code faster without making the design easier to understand. EGM puts a documented plan and separate review steps between an idea and a change to files. Google says thousands of its employees use the plugin daily to ship production code; that is the company’s account, not an independent evaluation of how well the checks catch defects.
There is also a useful boundary in the design. Security flaws are not specifically on the reviewers’ checklist, although teams can customise the reviewer roles. A design review is not automatically a security review, however much we might wish the labels did the work for us.
Our read
The promising part is not that several AI sessions are guaranteed to spot every mistake. It is the requirement to write down the plan, make another session explain it, and resolve objections before coding starts. That makes the work more inspectable for people, too. Developers can try the free plugin in Antigravity, but should add explicit security checks if that is what the project needs, and judge the workflow by the issues it actually catches.
What to watch
- When Google makes the plugin available in development environments beyond Antigravity.
- Whether Google or outside developers publish evidence about defects caught or missed.
- How teams customise reviewer roles, particularly for security and other specialist checks.
Discussion spark: Would you trust AI reviewers more when they start without the design conversation, or does that make them more likely to miss the context that matters?
Sources and evidence
- Source update (26 October 2026, 04:25 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.