How Thousands of Google Engineers Rethink Coding With the Elephant‑Goldfish Model

Photo by Karollyne Videira Hubert on Unsplash

A hidden bug caught by AI

The first public demo of the workflow showed a drift‑report command that compared the generated design against the actual code and flagged a bug where sold products could still be deleted instead of merely hidden. The bug surfaced only after the Goldfish reviewers ran their final check, proving that a fresh, context‑free eye can expose assumptions the original designer missed. The drift report’s detection aligned with the model’s goal of catching design flaws before implementation.

Why a spec‑first workflow spreads

Google reports that thousands of its engineers run the Elephant‑Goldfish Model plugin every day to ship production code. The same claim appears in the open‑source repository’s README, which repeats that daily users rely on the tool to draft specifications before any code is generated. The plugin runs three phases—brainstorm, draft, and review—each triggered by a command in the Antigravity environment. In the review phase, stateless Goldfish agents validate the specification without any prior conversation context, a pattern designed to surface blind spots early.

From internal experiment to open source

Google originally built the Elephant‑Goldfish Model as an internal spec‑driven planning workflow for its own coding agents, such as Claude Code and Codex. The internal tool was later open‑sourced under the Apache 2.0 license, with the repository published on GitHub under the google/elephant‑goldfish‑model organization. The open‑source release includes Markdown rules, defined sub‑agent personas, and ready‑made commands that orchestrate the Elephant and Goldfish interactions. The announcement blog post framed the release as a way to give external developers confidence by showing a method that has already been used to ship production code at Google.

Sources

1. GitHub - google/elephant-goldfish-model: Open-source skills for designing and reviewing software with AI agents · GitHub — “Thousands of Google engineers use EGM, an open-source plugin, to draft specs with deep-context agents and stress-test them with zero-context subagents before writing code. Elephant holds your project context and drafts the specification.”
2. EGM: Google's New Way To Plan With Claude Code Is Insane — “Google built a spec-driven planning workflow called the Elephant Goldfish Model (EGM) for coding agents like Claude Code and Codex. An 'elephant' agent remembers the full project context and writes a detailed plan, while context-free 'goldfish' sub-agents review that plan from multiple angles without seeing the original chat, so nothing said earlier can sway them. The review runs five rounds (numbered 5 to 9), producing a plan saved in a docs/designs folder before any code is built.”
3. Google open sources its Elephant-Goldfish Model plugin for AI coding - BetaNews — “Thousands of Googlers use this plugin every day to ship production Google code.”
4. Google open-sources the Elephant-Goldfish Model plugin its engineers use to catch AI coding mistakes - Help Net Security — “Google says thousands of Googlers use the plugin every day to ship production code.”