What is Elephant-Goldfish Model?
The Elephant-Goldfish Model (EGM) is an open-source coding-agent plugin from Google, released under Apache 2.0. Its slogan is 'design with an elephant, review with a goldfish': the Elephant is a deep-context agent that holds the project and drafts a specification, while Goldfish reviewers are fresh zero-context subagents that read only the draft and its cited sources, so they catch unstated assumptions the main agent carried over. You make the trade-offs and approve the design before any code is written.
It ships seven skills, egm-brainstorm, egm-draft, egm-resolve, egm-bootstrap, egm-review, egm-review-loop and egm-drift-report, and installs as a plugin for Antigravity, Claude Code, Codex, Claude Cowork and ChatGPT Work. The workflow runs /egm-brainstorm into a concepts brief, /egm-draft into a design draft, then /egm-review, with approved specs saved under docs/designs.
The README says thousands of Google engineers use it, though it also states this is not an officially supported Google product. The repo has about 6,500 stars; its first tagged release, EGM v0.1.0, shipped on 27 October 2026.
Is Elephant-Goldfish Model worth it?
Most agent review tools ask the same model that wrote a design to critique it, which mostly produces agreement. EGM's insight is to separate the roles by context. The Elephant drafts with everything it knows about the project; the Goldfish reviewers are fresh subagents handed only the document and its referenced sources, so they cannot lean on the conversation that produced the design and tend to surface the assumptions it left implicit. That the pattern comes from Google engineers, who say thousands of them use it, and carries about 6,500 stars within weeks of going public, is the strongest signal here.
The skills map to a real workflow rather than one monolithic command. egm-brainstorm pressure-tests an idea into a concepts brief, egm-draft turns it into a design draft a section at a time with pauses for your feedback, egm-review runs the fresh reviewers over the draft, and egm-drift-report compares an approved design with the checked-in code later. The project is explicit that EGM is for features that need design decisions, and that a small clear bug fix should use your normal workflow, which is the right line to draw.
The hard dependency is subagent support. Goldfish reviews require the harness to create fresh, isolated contexts, and the README says that if the agent cannot, EGM stops the review rather than fake it. Support therefore varies by harness, and the docs track live quirks: on Antigravity's agy CLI an unprefixed command behaved differently between versions 1.4.8 and 1.4.10, so the docs tell you to keep the /egm: prefix. Running it inside Codex means picking the skill from a picker instead.
Two caveats temper the enthusiasm. It is early software, a v0.1.0 first release with about 59 open issues, so the rough edges are in the harness plumbing as much as the prompts. And the disclaimer that it is not an officially supported Google product means there is no support contract behind the Google name, however many Google engineers rely on it internally. The method is sound and unusually well-argued; the packaging is a few months old.
Where it earns its place
A backend engineer adding webhook delivery to a service in Claude Code
Settle retry and duplicate-event behaviour before writing any code
/egm-brainstorm the feature, /egm-draft the design, then /egm-review so fresh reviewers challenge the retry and idempotency assumptions. Free; the cost is the extra subagent runs on your existing Claude plan and the time to resolve findings before implementing.
A platform team standardising agent-assisted design across several products
Give every feature a reviewed design document in the repo
Install the plugin across Antigravity and Codex users, save approved specs under docs/designs, and run egm-drift-report later to compare code against the agreed design. Confirm each harness can create subagents first, since EGM halts reviews where it cannot.
A tech lead evaluating EGM against spec-kit
Decide whether zero-context review is worth the subagent requirement
Run both on one feature: spec-kit's documents go through pull requests, while EGM adds adversarial review but depends on subagent support. Weigh a v0.1.0 project against spec-kit's larger, longer track record.
Before you commit
- Setup
- Install as a plugin: claude plugin marketplace add google/elephant-goldfish-model then claude plugin install egm@egm; agy plugin install for Antigravity, codex plugin for Codex. Start a new session, then run /egm-brainstorm; the first design takes a brainstorm, draft and review pass.
- Integrations
- Antigravity (app, IDEs and agy CLI), Claude Code, Codex, Claude Cowork and ChatGPT Work; seven egm- skills invoked by slash command or skill picker. Reviewers run as fresh subagents.
- Data and privacy
- Runs through your coding agent's own model provider and writes Markdown design documents into your repo; nothing is sent to a Google service by the plugin. Apache 2.0; the README states it is not an officially supported Google product.
- Cost at scale
- No paid edition. The cost is extra model usage on your existing plan, since brainstorm and review spin up subagents, plus the reviewer time to resolve findings before implementation.
- Limits
- Goldfish reviews need subagent support and stop where it is missing; about 59 open issues; harness-specific quirks (agy CLI 1.4.8 vs 1.4.10 command prefixing); first release is v0.1.0.
What’s good
- Zero-context reviewers expose assumptions the drafting agent carried over
- Seven skills cover brainstorm, draft, review and drift detection
- Installs for Antigravity, Claude Code, Codex, Claude Cowork and ChatGPT Work
- Stops rather than fakes a review when subagents are unavailable
What’s not
- Needs a harness that can spawn fresh subagents; support varies
- First release (v0.1.0) with about 59 open issues
- Not an officially supported Google product, despite the name
- Overhead only suits features with real design decisions
Best for
Teams using Antigravity, Claude Code or Codex that want a design reviewed by something other than the agent that wrote it.
Elephant-Goldfish Model alternatives
Elephant-Goldfish Model: common questions
- Is Elephant-Goldfish Model worth it?
- EGM's zero-context reviewers are the cleverest answer yet to an agent rubber-stamping its own design, as long as your harness can spawn the subagents it needs. We rate it 4.2 out of 5.
- How much does Elephant-Goldfish Model cost?
- Free and open source (Apache 2.0).
- What are the downsides of Elephant-Goldfish Model?
- Needs a harness that can spawn fresh subagents; support varies; first release (v0.1.0) with about 59 open issues; not an officially supported Google product, despite the name; overhead only suits features with real design decisions.
- Who is Elephant-Goldfish Model best for?
- Teams using Antigravity, Claude Code or Codex that want a design reviewed by something other than the agent that wrote it.
- How much does Elephant-Goldfish Model cost for a team or in production?
- No paid edition. The cost is extra model usage on your existing plan, since brainstorm and review spin up subagents, plus the reviewer time to resolve findings before implementation.
- Is Elephant-Goldfish Model safe for company data?
- Runs through your coding agent's own model provider and writes Markdown design documents into your repo; nothing is sent to a Google service by the plugin. Apache 2.0; the README states it is not an officially supported Google product.
- What are the best Elephant-Goldfish Model alternatives?
- The closest alternatives we have reviewed are Spec Kit, pdd-kit and Superpowers.
Answers checked Oct 30, 2026.
Make Elephant-Goldfish Model? Show this rating on your site
