Google open-sources the Elephant-Goldfish Model: design with an elephant, review with a goldfish

Google has published the Elephant-Goldfish Model (EGM) — an open-source set of skills that lets thousands of its engineers use AI agents to draft software designs and then have a team of "memory-free" reviewers examine them from zero. The goal is specific: end the pattern where an AI writes a spec, reviews its own spec, and inevitably praises it — letting unstated assumptions and missing edge cases slip through.

Quick answer

• The Elephant-Goldfish Model (EGM) is an open-source Google plugin (Apache 2.0) containing skills for designing and reviewing software with AI agents; the GitHub repo google/elephant-goldfish-model was created on 13 September 2026 and passed 10,000 stars as of this article's check (2 Nov 2026).
• The core idea: "Design with an elephant, review with a goldfish" — the "elephant" agent holds the full project context and drafts the spec; "goldfish" agents are brand-new subagents with no conversation history that read only the draft and its referenced sources to catch hidden assumptions and blind spots.
• The human keeps the decision rights: you resolve the architectural trade-offs and approve the spec before any agent writes code.
• The toolkit has 7 skills, usable together or individually: egm-brainstorm, egm-draft, egm-review, egm-review-loop, egm-resolve, egm-bootstrap and egm-drift-report.
• Installs on multiple agents: Claude Code, Antigravity (agy), OpenAI Codex, plus dedicated setup guides for Claude Cowork and ChatGPT Work.

What is the Elephant-Goldfish Model?

The Elephant-Goldfish Model (EGM) is an open-source set of Google skills for designing and reviewing software designs together with AI agents, before code is written. In this model, one agent plays the "elephant" — remembering the entire project context, requirements and earlier decisions — to explore the problem and draft the design specification. Then "goldfish" agents — completely fresh subagents that know nothing of the earlier conversation — read only the draft and the source material it references, and point out unstated assumptions, unclear requirements and implementation risks. The user decides the trade-offs, approves the final design and chooses what gets built.

According to the project's README, thousands of Google engineers use EGM in their daily work. The project's home is egm.dev, and the underlying idea comes from Dave Rensin's essay on elephants, goldfish and a new golden age of software engineering.

The problem it solves: AI grading its own homework

When one AI agent both writes a spec and reviews that same spec within a single context, it carries every assumption it formed while writing — and almost always concludes the draft is "fine". Anyone who has vibe-coded a serious project knows the result: code full of hidden assumptions, missing edge cases, and problems that only surface deep into implementation.

EGM attacks exactly this by separating memory from review: the writer remembers deeply (elephant), the reviewers forget everything (goldfish). Knowing nothing of the "original intent", the goldfish is forced to read the draft as an outsider — precisely the viewpoint that catches the logic holes its author, human or AI, is blind to.

The three roles

Role	Context it sees	Responsibility
Elephant	Your project, requirements and earlier decisions	Explore the problem and draft the specification
Goldfish	The document and its referenced source material, without the earlier conversation	Find missing assumptions, unclear requirements and implementation risks
You	The goals and constraints that matter	Resolve trade-offs, approve the design and decide what to build

The 7 skills in the EGM plugin

Skill	What it does
egm-brainstorm	Tests an idea, discusses requirements and trade-offs; saves the outcome as concepts_brief.md, with fresh reviewers challenging the proposal already at this stage
egm-draft	Turns a brief or existing notes into a design_draft.md design document, pausing after each section for your feedback
egm-review	Reviews the draft with fresh agents across five stages: clarity, fit with the codebase, and the implementation plan
egm-review-loop	Runs a configured group of reviewers for a specific review task
egm-resolve	Works through the review findings and agrees the changes
egm-bootstrap	Creates or updates documentation for an existing codebase
egm-drift-report	Compares checked-in documents with the current source code to detect drift

Every skill can be used on its own; the full sequence is not mandatory. In Claude Code and Antigravity the commands carry the /egm: prefix, e.g. /egm:egm-brainstorm.

The standard workflow: from idea to approved spec

1. Brainstorm: run /egm:egm-brainstorm with a feature description (the project's canonical example: "Add webhook delivery to this project" — including questions like what happens when a delivery fails, or the same event arrives twice). The agent discusses requirements and trade-offs, fresh reviewers challenge the proposal, and the agreed direction lands in concepts_brief.md.
2. Draft: /egm:egm-draft turns the brief into a detailed design_draft.md, pausing after each section for your feedback right in the chat.
3. Review: /egm:egm-review design_draft.md lets the goldfish examine the draft. You choose which findings to address, working through them with egm-resolve if needed.
4. Approve and implement: once you approve, the accepted spec is saved under docs/designs/ in the repo as the reference document — and only then does the agent implement it.

How to install

Install the complete plugin (so the skills can read the bundled reviewer prompts and document standards), then start a new session:

Claude Code:

claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

Antigravity CLI (agy):

agy plugin install https://github.com/google/elephant-goldfish-model

OpenAI Codex:

codex plugin marketplace add google/elephant-goldfish-model
codex plugin add egm@egm

Claude Cowork and ChatGPT Work have their own setup instructions in the repo's docs/harnesses.md.

When to use it — and the limits to know

• Use it when a feature needs design decisions before code: a new mechanism, an architecture change, an external integration. That is where a well-reviewed spec saves many rounds of code fixing later.
• Skip it for a small, clear bug fix — the project itself recommends your normal coding workflow there.
• Technical condition: goldfish reviews require the agent to create fresh-context subagents; if the environment cannot, EGM stops the review instead of pretending one happened.
• Not a replacement for tests or human judgment: the project documentation states plainly that an EGM review does not replace tests or an engineer's judgment.
• Legal note: EGM is released under Apache 2.0, but the repo also states it is "not an officially supported Google product" — worth considering before adopting it in a company's process.

Frequently asked questions

Is the Elephant-Goldfish Model a new AI model?

No. EGM is not a language model; it is a plugin of packaged skills (workflows) and reviewer prompts that runs on the coding agents you already use, such as Claude Code, Antigravity or Codex. "Elephant" and "goldfish" are two agent roles, not two models.

Does using EGM cost extra?

The plugin is free under Apache 2.0. The real cost is the tokens your coding agent consumes running extra brainstorm, draft and especially goldfish-review rounds — in exchange for a stronger spec and fewer code-fix cycles.

How is this different from opening a new chat and asking the AI to "review this spec"?

It is systematic rather than ad hoc: EGM ships dedicated reviewer prompts, a five-stage review process (clarity, code fit, implementation plan…), a consistent findings format and a disciplined resolve/approve step, instead of one generic request with unpredictable results.

Which agents does EGM support?

The repo officially supports Claude Code, Antigravity (CLI, app and IDE editions) and OpenAI Codex, with setup guides for Claude Cowork and ChatGPT Work. The common requirement: the environment must be able to create fresh-context subagents for the goldfish step.

Why a "goldfish"?

Goldfish are famous for short memories — a metaphor for a reviewer deliberately given no prior conversation context. Remembering nothing forces the goldfish to read the draft as an outsider and catch the hidden assumptions the deep-remembering elephant can no longer see.

Conclusion

The Elephant-Goldfish Model shows a maturing direction for AI-assisted programming: instead of trusting one agent to perform and grade itself, separate the writing and reviewing roles with a hard context boundary, and keep the human in the final approval seat. For any team using Claude Code, Antigravity or Codex on serious projects, EGM is an open-source workflow worth trying — especially since a wrong spec always costs more than a careful review round.

Sources: the GitHub repo google/elephant-goldfish-model (README and official documentation, verified 2 Nov 2026) and the introduction post by the Code Web Khong Kho page on Facebook, edited by Wotech Electronics.