A Google plugin brings in AI reviewers before coding begins
Elephant-Goldfish Model has one AI agent draft a feature plan, then sends it to other agents for review in fresh sessions.
An online store wants to notify its warehouse every time a customer places an order. The plan calls for an automated message. But if the warehouse receives the same message twice, could it ship two packages? And if the message never arrives, who resends it? The Elephant-Goldfish Model guide starts with this kind of problem. A developer and an AI assistant can settle a question in conversation and still leave the answer out of the document that guides the coding. One AI agent drafts the plan, others review it
The first agent, nicknamed Elephant, knows the project and follows the discussion with the developer. It turns their decisions into a document that explains what the feature should do and how it fits into the software. In the online store example, the plan might specify that every order has a unique number and that the warehouse uses it to recognize a duplicate message.
The reviewers, nicknamed Goldfish, each start in a fresh session. They read the plan and, depending on what they’re checking, the relevant project files. The earlier conversation stays in the original session. In one step of the review, an agent explains in its own words what the software is supposed to do. If the plan simply says “handle duplicates,” the reviewer can ask how the warehouse is supposed to tell whether two messages refer to the same order. Other reviewers then check whether the files and functions named in the plan exist in the project.
The developer reviews the feedback, decides what to change, and approves the plan. They can then ask their assistant to write the code using the approved version. Where the plugin works
Originally developed for Google’s own teams, the plugin is available under the Apache 2.0 license. The compatibility guide documents tests in the Antigravity command-line tool, Claude Code, and the Codex command-line tool. The plugin is also available in the ChatGPT and Codex desktop apps. The review requires an app that can launch agents in fresh sessions. If that capability is unavailable, the plugin stops the review.