Google open-sources the Elephant-Goldfish Model plugin, letting AI separately handle the writing and verification of design documents

On October 22, 2026, Google announced the open-source release of the Elephant-Goldfish Model (EGM) plugin, which supports design and review before implementation in AI-assisted software development. It combines an AI that creates design documents based on dialogue with the user and an AI that verifies the documents without inheriting the history of that dialogue, thereby surfacing ambiguities and oversights in specifications.

Introducing the Elephant-Goldfish Model plugin: scaling design-driven AI software development - Google Open Source Blog

google/elephant-goldfish-model - GitHub

The plugin was developed internally for Google's software engineers. According to the company, thousands of employees use this plugin every day to develop Google's production code. The public repository includes customizable Markdown rules, role settings for review sub-agents, review procedures, and more. The license is Apache 2.0.

Elephant advances the design, Goldfish verifies the design document

Google points out that while AI speeds up writing code, it also accelerates the generation of bugs and low-quality code. It also cites the problem of losing prior context within a session with AI. With EGM, so that people can understand the system's structure and the reasons behind design decisions, the design is worked out with the AI before writing code, and those decisions are recorded in a document. According to the official blog, the aim of this approach is to get AI to generate code as intended from the first attempt.

According to an explanation by Google's Dave Rensin cited in the official blog, the name Elephant comes from the English expression "An elephant never forgets." Elephant acts as the lead AI, conversing with the user and advancing the design. It retains the project structure, requirements, and decisions made so far within the session, and examines the problems to be solved and design options. At this stage it does not write code; instead, it compiles the decisions into a design document.

On the other hand, Rensin likens Goldfish to a new AI session that does not inherit past exchanges and reads only the information given to it. Goldfish verifies the design document without being bound by the same assumptions as the designer.

In the plugin, the lead agent follows the review skill's procedure and launches Goldfish as a sub-agent that runs in an independent context. Depending on the review stage, there may be a single Goldfish checking, or multiple Goldfish verifying from different perspectives. Each Goldfish reads the design document and the materials referenced therein, looking for unstated assumptions, unclear requirements, and implementation risks. The lead Elephant revises the design based on the returned review results.

From conception to documentation to verification, then to implementation

The official blog introduces the flow of collaborative design using the three skills included in the plugin. The plugin automates the flow of passing design documents and review results between Elephant and Goldfish and repeatedly verifying the design.

The first, egm-brainstorm, has Elephant read the structure of the development target and prior design decisions, discuss requirements and trade-offs, and question assumptions. The next, egm-draft, organizes the problem definition, technical approach, and alternatives considered, and compiles them into a Markdown design document along with an implementation plan listing the files to be changed. The content is refined while incorporating user feedback.

In egm-review, Goldfish verifies the design. The checks cited in the official blog are the following three.

Comprehension check: Goldfish, having read the design document, explains the system's structure and design intent to Elephant. If it cannot explain accurately, the ambiguous parts of the design document need to be rewritten.

Critical review: Multiple Goldfish look for exceptional cases, hard-to-see dependencies, and gaps in logic. Elephant, having received the findings, revises the design.

Confirmation of readiness for implementation: It checks whether the information needed for implementation—such as file paths, data structures, and function arguments and return values—is complete and whether any ambiguity remains.

The workflow guide shows a flow in which the user selects the items to address based on Goldfish's findings and has another person review the design as well. After the user approves the design, implementation is requested from a coding agent.

Each skill can be used individually as needed. The guide states that it is not necessary to go through the entire design process for small, clearly scoped bug fixes.

First tagged release v0.1.0 published on October 27

On October 27, the first tagged release, v0.1.0, was published. It contains a total of seven skills, including the three introduced in the official blog plus the following four.

egm-review-loop: Manages the flow of document review by multiple Goldfish and the revision and re-review following findings. When applying revisions or proceeding to the next review, user confirmation is required.

egm-resolve: Elephant considers Goldfish's findings and presents proposed revisions or the rationale for rebuttals. Revisions approved by the user are reflected in the document.

egm-bootstrap: Examines existing code and creates or updates a README for each directory. It documents each part's role, interfaces, conditions to be observed, and so on.

egm-drift-report: Compares design documents and READMEs with the actual code and reports discrepancies such as unimplemented features or processing that differs from the design.

The distribution ZIPs for Claude and for OpenAI include the seven skills that make up the plugin, instructions for reviewers, document creation standards, and more. As of October 28, the README provides installation instructions for Antigravity, Claude Code, and Codex. Detailed procedures and verification status for each environment are compiled in the setup guide.

However, the release notes state that verification of operation in each runtime environment using these distribution ZIPs is incomplete, and that it is not yet at a stage that can be called stable. Also, in environments where Goldfish cannot be launched in an independent context, review is stopped.

The README also clearly states that review is not a substitute for testing or human judgment, and that this is not an officially supported Google product.