Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, a design-and-review system designed to force AI coding agents to write specifications before producing code. Under the Apache 2.0-licensed framework, a long-session subagent called the Elephant drafts a project spec without writing code, while stateless subagents known as Goldfish independently review the spec with zero conversation history. The goal is to catch flawed assumptions in Markdown files prior to pull request generation.

According to Google, thousands of its internal engineers already use the method to prevent session context collapse and AI-generated code slop. EGM functions as a set of reviewer personas, document standards, and skills rather than a new underlying language model. It features a five-stage review process that ranks findings by severity, from Blocker to Nit.

The framework supports multiple developer tools including Antigravity CLI, Claude Code, and Codex, with automated end-to-end fixtures already operational across several environments. Google notes that EGM is an open-source community effort rather than an officially supported product.

Why it matters

  • Eliminates context drift by forcing AI coding agents to separate specification design from code implementation.

  • Reduces pull request review overhead through automated, zero-context subagent audits before code generation.

  • Provides an Apache 2.0 framework compatible with popular harnesses like Claude Code, Codex, and Antigravity.

Source: marktechpost.com