Free guide

Google's Elephant-Goldfish Model: Fresh-Eyes Spec Review for Coding Agents

A free, open-source plugin that has one agent write the plan with you and fresh agents with no memory of your chat review it, before any code gets written. What it is, how it works, how to set it up, and the fine print.

By Beau Johnson·
The Short this guide goes with: Google's Elephant-Goldfish Model: Fresh-Eyes Spec Review for Coding Agents

Quick answer

The Elephant-Goldfish Model (EGM) is an open-source plugin from Google (Apache 2.0) for Claude Code, Codex, Antigravity and ChatGPT. An elephant agent holds your project context and drafts a design spec with you; goldfish reviewers are fresh subagents that read only the draft and its source material, without the earlier conversation, to expose unstated assumptions. You make the trade-offs and approve the spec before your agent builds it.

The short version

  • EGM is Google's open-source plugin (Apache 2.0) for designing and reviewing software with agents. Google says thousands of its engineers use it.
  • The elephant knows your project and drafts the spec with you. The goldfish are fresh subagents that never saw your chat, so they read the plan with fresh eyes.
  • It installs in a minute in Claude Code, Codex or Antigravity. Reviews need an agent that can start fresh subagents.

How it actually works

You can't proofread your own essay

When you reread your own writing, your brain fills in what you meant. A coding agent that has been in a long chat with you has the same problem: the more it knows, the easier it is to skip over what it is assuming. EGM splits the job so the one who writes the plan is not the only one who checks it.

The elephant: full context

The elephant is your main agent session. It holds your project, requirements and earlier decisions, brainstorms the idea with you, and drafts the specification section by section, pausing after each section for your feedback.

The goldfish: zero context

Goldfish reviewers are fresh subagents. They examine the draft and the source material it references, without the earlier conversation, to find missing assumptions, unclear requirements and implementation risks. The review runs in five stages: initial zoom out, comprehension, design critique, intermediate zoom out and implementation readiness.

You decide

You choose which findings to address, make the architectural trade-offs and approve the spec. Then you ask your coding agent to build it. After it is built, a drift report can compare the code with the spec.

How to set it up

  1. 1.Install it in Claude Code

    Add the marketplace, install the plugin, then start a new session.

    claude plugin marketplace add google/elephant-goldfish-model
    claude plugin install egm@egm
  2. 2.Or install it in Codex or Antigravity

    Codex CLI uses the first two lines; Antigravity CLI uses the last one. In the ChatGPT and Codex desktop apps, install the published EGM plugin from the plugin directory.

    codex plugin marketplace add google/elephant-goldfish-model
    codex plugin add egm@egm
    
    agy plugin install https://github.com/google/elephant-goldfish-model
  3. 3.Brainstorm and draft one real feature

    Pick a feature that needs design decisions (the README's example is webhook delivery: what should happen when a delivery fails, or the same event arrives twice?). In Claude Code and Antigravity add the /egm: prefix; in Codex pick the EGM skill from the skill picker. Brainstorm saves concepts_brief.md; draft writes design_draft.md.

    /egm:egm-brainstorm Add webhook delivery to this project.
    /egm:egm-draft
  4. 4.Send in the goldfish

    Run the review on the draft path the agent gave you. Fresh reviewers check the design, its clarity, its fit with the code and its implementation plan. Pick the findings that matter, then approve and build.

    /egm:egm-review design_draft.md

The fine print

  • Apache 2.0 licensed and free. The README says: "This is not an officially supported Google product."
  • Goldfish reviews need fresh subagent contexts. If your agent can't create them, EGM stops the review instead of faking it in the same chat.
  • The README says a review does not replace tests or human judgment. You still test the code and approve the design.
  • By design it gives a lot of feedback. The getting-started guide says to take what is relevant and ignore the rest.
  • The idea comes from Dave Rensin's essay "Elephants, Goldfish, and the New Golden Age of Software Engineering".

FAQ

What is the Elephant-Goldfish Model?

An open-source plugin from Google for designing and reviewing software with AI agents. An elephant agent with full project context drafts the spec with you, and goldfish subagents with no memory of the conversation review it before any code is written.

Is the Elephant-Goldfish Model free?

Yes. It is open source under the Apache 2.0 license. You still pay for the coding agent you run it in, like your Claude or ChatGPT plan.

Which coding agents does EGM work with?

Claude Code, Claude Cowork, Codex (CLI and desktop), ChatGPT Work and desktop, and Google Antigravity (CLI, app and the VS Code extension). Reviews need an agent that can start fresh subagents.

Why use a goldfish instead of asking the same agent to review its plan?

The agent that wrote the plan carries the whole conversation, so it reads what it meant. A goldfish reviewer only sees the draft and its sources, so unstated assumptions and gaps are easier to spot.

Does EGM replace tests or code review?

No. The README says a review does not replace tests or human judgment, and the getting-started guide recommends a human review of the design too.

Sources

Facts checked 2026-11-09 against the EGM README, getting-started guide, the egm-review skill, egm.dev and the GitHub API (Apache 2.0, v0.1.1 released 2026-11-06).

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides