--- title: "How to Install and Use Google's EGM Plugin (Elephant Goldfish)" url: "https://goremy.ai/blog/articles/google-egm-elephant-goldfish-model-plugin" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-31T09:40:06.695Z" updated: "2026-10-31T09:40:06.695Z" topic: Agent Harnesses and Skills description: "Google's EGM plugin has an agent write a design doc, then fresh sub-agents review it. Install steps for Claude Code, Codex and Antigravity, plus commands." --- # How to Install and Use Google's EGM Plugin (Elephant Goldfish) > Google's EGM plugin has an agent write a design doc, then fresh sub-agents review it. Install steps for Claude Code, Codex and Antigravity, plus commands. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-31 ## What is Google's Elephant Goldfish Model (EGM)? The Elephant Goldfish Model, or EGM, is an open-source plugin from a Google team that makes your coding agent design a feature before it writes code. One long-memory agent (the elephant) writes a design document. Then fresh sub-agents with no memory of your conversation (the goldfish) review it. You make the final calls. It is Apache 2.0 licensed and runs in Antigravity, Claude Code and Codex. ## TL;DR - EGM splits work into three roles: an **elephant agent that remembers everything**, **goldfish sub-agents that start cold**, and you as the decision maker. - Installation is a two-step marketplace flow in Claude Code and Codex, and a single plugin install from a GitHub URL in Antigravity. - Install the **whole plugin, not just the skills folder**, because the skills read reviewer personas and document standards at runtime. - The main flow is brainstorm, draft, review, then resolve, all under an `/egm` command prefix. - If your agent **cannot spawn fresh sub-agents, EGM stops the review** instead of faking it in the same context. - Reviews are token-hungry, and the repo itself says EGM is not worth it for small, clear bug fixes. ## Where does the idea come from? EGM builds on an essay titled "Elephants, Goldfish, and the New Golden Age of Software Engineering." The problem it targets is familiar to anyone who has spent an hour with an agent on one feature. You and the agent pile up shared assumptions. The design doc then reads fine to both of you, because you both know what is missing. A fresh reader does not. When a goldfish asks what happens if a webhook delivery fails, or what a term even means, it finds the gaps that would otherwise surface during implementation. The repo carries a caveat: it is not an officially supported Google product. Treat it as a project from a Google team, not a company launch. A video review of the project noted roughly 2,600 GitHub stars after about six weeks. ## How do I install EGM? Commands below follow the project's readme as described by a creator walkthrough. Check the repo for the exact repository path before running them. **Claude Code** ```bash claude plugin marketplace add google/elephant-goldfish-model claude plugin install egm@egm ``` **Codex** ```bash codex plugin marketplace add google/elephant-goldfish-model codex plugin add egm@egm ``` **Antigravity** Antigravity uses a single command with the GitHub URL of the repo, in the form `ag plugin install `. After installing, start a new session. Install the full plugin rather than copying skill files by hand. The skills load reviewer personas and document standards while they run, so a partial copy breaks. The readme lists first-class install instructions for other Claude and ChatGPT surfaces too. The repo has a dedicated Claude Code folder with its own sub-agent adapters, and a separate Codex manifest. ## How do I use EGM day to day? All commands share the `/egm` prefix. The exact spelling of each command can differ by harness, so confirm in the readme. The workflow has four main steps. | Step | Command | What happens | |---|---|---| | 1 | brainstorm | You describe a feature. The agent asks clarifying questions, then sends six goldfish reviewers in parallel. Output is a concepts brief file. | | 2 | draft | The brief becomes a design document, written section by section with a pause for your feedback after each. | | 3 | review | Five review stages run against the draft. | | 4 | resolve | You go through the findings and agree on what to change. | Once you approve the design, you tell your agent to build it. ### What does the brainstorm step do? You give it something like "add webhook delivery to this project." Six goldfish look at the idea through different lenses. One is a contrarian that asks why the feature should exist at all and whether an existing library already covers most of it. Another looks at cost, quota and token economics. You end up with a brief that records the direction you agreed on. ### What are the five review stages? 1. **Zoom out.** A skeptical architect hunts for complexity that is not needed. 2. **Comprehension.** A naive reader flags every term and acronym left unexplained. 3. **Hostile design critic.** This reviewer examines architecture, scaling and risks. 4. **Zoom out again.** A second pass at the big picture. 5. **Implementation readiness.** A pragmatic tech lead checks that no design decision is left for whoever implements the plan. The project warns that this produces a lot of feedback by design. Take what is relevant and ignore the rest. ### What other commands exist? - **bootstrap** walks an existing codebase from the bottom up and writes readme files for each folder, then summarizes them toward the root. This gives the elephant a map of the project. - **drift report** compares your design docs against the current code and shows where they no longer match. The project recommends running it weekly. ## Why do the reviewers stay honest? Two design rules matter most. First, the review skill bans flattery. Reviewers may not call your idea great or brilliant. They lead with what is wrong or risky and offer the strongest counterargument to your proposal. Second, isolation is enforced. Each reviewer gets read-only tools and its own context. If the agent cannot create fresh sub-agents, EGM halts the review. A review run in the same context is not a goldfish review, and the plugin refuses to pretend otherwise. The repo also states that no harness gets a forked copy of the skill text. There is one set of skills and personas, and each harness maps the word "sub-agent" to its own tool. ## Can I run EGM in other agents like Cline? Probably, with more manual setup. Cline is not on Google's official list, but it supports `SKILL.md` skills and sub-agents, which is most of what EGM needs. The skills are plain markdown, and the reviewers are persona files in an agents folder. For Codex, EGM ships no named sub-agents at all. A general-purpose sub-agent reads the persona file first. Another harness can do the same. If you try this, clone the whole repo and point the agent at the skill you want. Then confirm the reviewers run as fresh sub-agents. Reviewers that share your chat defeat the purpose. One idea worth testing is mixing models. A goldfish running on a different model than the elephant avoids shared assumptions and shared blind spots. The creator who demonstrated this ran it on open models through a Cline subscription. That is one person's setup, not part of the plugin. ## Is EGM worth the tokens? It depends on the feature. The readme says that for a small, clear bug fix you should use your normal workflow. Its FAQ admits that design and review take time and may not pay off for small changes. Cost is real. One review can launch five or six reviewers, and each reads your document and the referenced code from scratch. Repeated review rounds on a low-tier subscription can hit usage limits quickly. Use EGM where decisions need to be made before code exists: new subsystems, integrations, anything with unclear scope. ## Frequently Asked Questions ### Does EGM only work with Google's tools? No. It works in Antigravity, Claude Code and Codex, with install instructions for each. Because it consists of skills and sub-agent prompts, other harnesses that support both can run it with some manual setup. ### Can I copy just the skills folder? No. The skills read reviewer personas and document standards at runtime. Copying only the skills breaks them, so install the whole plugin or clone the whole repo. ### What happens if my agent cannot spawn sub-agents? EGM stops the review. It will not simulate reviewers in the same context, since that would not give you a fresh reader. ### Is EGM free? The project is open source under the Apache 2.0 license. You still pay for whatever model usage the elephant and goldfish agents consume. ### Do I need EGM for every change? No. The project says to skip it for small, clear bug fixes. It is meant for features where design decisions come first.