Elephant-Goldfish Model: from idea to code with AI agents

A small group at Google has released the Elephant-Goldfish Model (EGM), an open-source, Apache-2.0-licensed workflow for building software with AI agents that is already used by thousands of Googlers. EGM guides developers from brainstorming a feature through drafting a design, independent review by agents that don't see the drafting conversation, resolving findings, and checking for drift between design and code. It ships as a Claude Code plugin with commands such as /egm:egm-brainstorm, /egm:egm-draft, /egm:egm-review and /egm:egm-drift-report, and also works with Codex and Antigravity; the project is built at Google but not officially supported by it.

I’ve been working with a small group at Google on Elephant-Goldfish Model https://egm.dev , an open-source workflow for building software with AI agents. Thousands of Googlers already use it. You can start with a feature you want to build. EGM helps you talk it through, write a design, get independent reviews, and work through the findings. As you build, you can check where the code has drifted from the design. The agent helping you draft keeps your project context. Separate reviewers read the design and relevant code without seeing that conversation. They look for missing decisions and assumptions. You choose what to change. In Claude Code, install the plugin: claude plugin marketplace add google/elephant-goldfish-model claude plugin install egm@egm Run /egm:egm-brainstorm with a feature you’re considering, then /egm:egm-draft to turn it into a design. When you have a draft, run /egm:egm-review docs/designs/my-feature.md . Some findings will be wrong. Check them against your intent and the code, then use /egm:egm-resolve to work through the ones that matter. /egm:egm-drift-report checks for differences between your design and implementation as things change. It also works with Codex and Antigravity. See egm.dev https://egm.dev for setup instructions. The plugin is Apache-2.0 licensed. Your usual coding-agent usage costs apply. Try it on a feature with a few real decisions to make and let me know how it goes in the comments. Built at Google, not officially supported by Google.