Skip to content
Neural Daily

ToolsNews

Google EGM: design first, then code for AI agents

🇭🇺 Olvasd el magyarul →

Google open-sourced the Elephant-Goldfish Model plugin, which thousands of its engineers use daily. The idea: the AI writes a detailed design first and code only after that.

By: Mikola Máté Patrik · · 2 min read

Elephant-Goldfish Model graphic with a blue elephant icon labelled Remembers everything and an orange goldfish icon labelled Remembers nothing
Image: Google · Press kit · Official image from the Google Open Source Blog announcement · Source

What launched?

On October 22, Google introduced the Elephant-Goldfish Model (EGM) plugin and released it under Apache 2.0 (Google Open Source Blog). According to the post, it started as an internal tool and thousands of Googlers now use it every day to ship production code. The premise: AI coding speeds up not only development but also the production of bugs and low-quality code. In the authors' words, we often end up auto-scaling our mistakes.

How it works

The method uses two roles. The Elephant is a long, context-rich AI session that holds the full architecture and every decision. No code is written there: the AI interviews the developer and they debate the options together. Goldfish are stateless, zero-context sessions that only see the finished spec. Because they share none of the designer's blind spots, they act as objective critics.

  • egm-brainstorm: design under a "No Code" rule, debating trade-offs and assumptions.
  • egm-draft: a Markdown design doc covering the problem, the technical plan, alternatives considered and the list of files to change.
  • egm-review: one Goldfish first explains the plan back, then several Goldfish act as skeptical reviewers hunting for gaps. A final Goldfish makes a go/no-go call on whether the spec is precise enough to code.

Where it runs

The plugin targets Google Antigravity first, but the GitHub repo also has install instructions for Claude Code and Codex, with separate docs for Claude Cowork and ChatGPT. The repo states it is not an officially supported Google product, so there is no product support behind it.

Why it matters

Most teams today ask for code straight from a prompt and fix the bugs afterwards. EGM flips that: it tries to catch most problems in the design before a single line of code exists. The memoryless reviewer idea is simple but effective, and it can be adopted in any agentic tool even without the plugin. If you build larger features with AI agents, try it on a pilot project.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles