The Elephant Goldfish model (EGM) is Google's open-source method for spec-driven development with coding agents. One long session with the full project context drafts a design document, and then fresh subagents that never saw that conversation review it before anyone writes code. Google announced the plugin on 22 October 2026. The repository passed 2,000 stars on 27 October, the same day it tagged its first release, v0.1.0. The release notes call it "not ready for stable-release claims".
What is the Elephant Goldfish model? #
The Elephant Goldfish model is a plugin with no executable code: seven skills, nine agent personas (seven of them reviewers) and five review rubrics, all in Markdown, licensed under Apache 2.0. It has no CLI and no service. It runs on the coding agent you already use, with that agent's model and permissions. The repository describes it as "an open-source method and skill set created at Google for spec-driven development with LLMs", and adds that it "is not an officially supported Google product".
The method is older than the plugin. Dave Rensin set it out in an April 2026 essay whose rule is "Feed the Elephant; test it against the Goldfish." EGM's own architecture document, dated June 2026, calls the plugin an implementation of spec-driven development based on that essay. In the eight-framework comparison I published on 12 October, spec-driven tools are the ones that write a proposal, a spec and a task list before an agent writes code. EGM covers the spec and stops there.
The elephant drafts the spec, the goldfish review it cold #
In the Elephant Goldfish model, the elephant is the long, context-heavy agent session that interviews you and writes the design, and the goldfish are stateless subagents that read only the draft and the source code it points to. Google's blog defines the elephant as a session that "holds the end-to-end architectural vision and project memory. No code is written here". The goldfish "are stateless, zero-context sessions used strictly to validate specifications."
A reviewer who missed the conversation can catch the assumptions it never wrote down. The README states each role's context: the elephant gets "your project, requirements, and earlier decisions", and a goldfish gets "the document and its referenced source material, without the earlier conversation". You make the trade-offs and approve the spec. Rensin's essay also counts the design document as part of the elephant, while the blog and the repository use the word only for the session.
The Elephant and Goldfish cards and the workflow on egm.dev. Image: egm.dev screenshot in google/elephant-goldfish-model by Google LLC and the EGM contributors, Apache 2.0, cropped.
One feature through EGM: brainstorm, draft, five review stages #
A feature goes through EGM in three commands, and the README walks through one that adds webhook delivery. In Claude Code, each command takes an /egm: prefix:
/egm:egm-brainstorm Add webhook delivery to this project.
/egm:egm-draft
/egm:egm-review design_draft.md
The brainstorm discusses requirements with you, sends six fresh reviewers at the idea (one lens is "Contrarian / YAGNI") and saves concepts_brief.md. The draft step writes design_draft.md one section at a time and waits for your feedback after each. The design doc template has goals, non-goals, user journeys in Given / When / Then form, alternatives considered and a detailed implementation plan that lists every file to create or change.
The review runs five stages in a fixed order: initial zoom out, comprehension, design critique, intermediate zoom out and implementation readiness. Every stage runs in fresh subagents, three in parallel for the design critique and the readiness check. EGM writes no code. Once you approve the design, any agent builds it, and /egm-drift-report, which the docs suggest running weekly, compares the code with the spec afterwards.
EGM next to Spec Kit, OpenSpec and a plain AGENTS.md #
EGM fills the same before-code slot as Spec Kit and OpenSpec, but it stops at an approved design, and its distinctive part is the review by fresh-context reviewers. Spec Kit and OpenSpec carry a change from spec to implementation. A plain AGENTS.md has no phases at all: it is a set of rules every session reads.
| EGM | Spec Kit | OpenSpec | Plain AGENTS.md | |
|---|---|---|---|---|
| Form | Plugin: skills, personas, rubrics | Python CLI and commands | CLI and skills | One Markdown file |
| Licence | Apache 2.0 | MIT | MIT | Open format |
| Latest, 27 Oct 2026 | v0.1.0, 27 Oct | v1.2.3, 22 Oct | v1.15.0, 24 Oct | n/a |
| Stars, 27 Oct 2026 | about 2,000 | about 143,000 | about 73,000 | n/a |
| Agents | 3 verified, 2 partial | 41 integrations | 39 tools in its docs | Most agents; Claude Code by default only without a CLAUDE.md |
| Writes | Concept brief, design doc, review reports | Spec, plan, tasks per feature | Change folder merged into living specs | Nothing per task |
| Before code | 5-stage review by fresh subagents | clarify, checklist, analyze | Your review, openspec validate |
Nothing |
| After code | Drift report | implement, converge | Advisory /opsx:verify |
Nothing |
The figures come from each project's GitHub page and docs, read on 27 October 2026. Spec Kit has an adversarial review only as an extension, according to a May 2026 comparison, and OpenSpec has none. That review is the step EGM adds. None of these tools stops an agent from implementing: EGM's one hard stop, an instruction in the skill rather than code, is that a review halts if the agent cannot start a fresh subagent. The quality gates post covers tools that do block.
EGM loads little at start and up to 16 reviewers per review #
The Elephant Goldfish model costs almost nothing in context when it is idle and a lot when a review runs. I counted the files in the v0.1.0 clone, dividing characters by four, which gives a rough estimate rather than a tokenizer count. In Claude Code, all seven skills are user-invoked only, so their descriptions stay out of context. A new session carries the nine persona descriptions, about 200 tokens.
Running a command adds the skill, the Claude Code notes every skill reads first and any template it needs to the main session: about 6,000 tokens for the brainstorm, 8,700 for the draft and 7,500 for the review, or about 11,400 counting the two helper skills the review reads along the way. Reading the harness guide the skills point to adds about 1,700 more. Each reviewer starts with 1,500 to 2,100 tokens of instructions plus the design doc and every source file it cites. The project's own test runs used 9 to 16 reviewers for a full review. For comparison, one Spec Kit command with the template it fills loads about 2,100 to 6,100 tokens by the same count. Nobody has published the total token cost of a full EGM run.
EGM runs in Antigravity, Claude Code and Codex today #
Three coding agents pass the project's automated tests for installation, reviewer isolation and the full brainstorm, draft and review workflow: Google Antigravity, Claude Code and Codex. Claude Code needs version 2.1.271 or later: older versions lack the setting that keeps the project's CLAUDE.md out of a reviewer's context, and the project does not count them as isolated.
Claude Cowork and ChatGPT Work install, but have open verification gaps (#105, #106). In Cowork's test run, a usage limit stopped the review in the fourth of its five stages. Cursor, GitHub Copilot, Gemini CLI and OpenCode are not on the list. The plugin also has to be installed whole: harnesses.md warns that copying only the skills/ folder breaks it.
What Google has shown and what it only states #
Google has shown that EGM's workflow completes on fixture designs on each supported agent and that its reviewers do not see the parent conversation; Claude Code passed after a fix, loaded from a local folder rather than the marketplace. It has not published a measurement of whether specs get better. The testing record covers installation, an isolation probe and fixture runs dated 22 October. The repository also holds 23 behavioural test scenarios, in seven files, that check the skills follow their own steps, with no results published.
The rest is stated without data. "Thousands of Google engineers use EGM" comes with no count or date range. The blog says goldfish "share no 'blind spots' with the designer", but by default the Claude Code reviewers run on the same model as the elephant, so they can share the model's own blind spots. Digital Applied suggests running them on a different model. Rensin's figure that about 30% of a critic's suggestions are highly valuable is, in his words, anecdotal. His figure of "more than 100,000 Googlers" describes an internal tool he built with the method, not EGM. The FAQ hedges: "Completing the EGM workflow alone does not demonstrate a productivity improvement."
Even the count of review steps varies: three checks in the blog, four in the architecture document, five in the shipped skill.
Who needs EGM and who can skip it #
EGM fits a team that already writes design documents and wants each one attacked before an agent builds from it. The README says: "Use it when a feature needs design decisions before implementation." For "a small, clear bug fix", it says to use your normal coding workflow.
Skip it if you use Cursor, Copilot, Gemini CLI or OpenCode, if you want a tool that takes the spec through to merged code, or if a full review run would strain your token budget.
On Stackness, as of 27 October 2026, Claude Code is on 8 public stacks and Codex on 2, bots removed and mine included (how I count). None lists Spec Kit, OpenSpec or AGENTS.md yet, so these numbers are small. The AI tools developers list on Stackness show which agents people run today.
Where to start with the Elephant Goldfish model #
Start with a design you already have rather than a new feature. In Claude Code 2.1.271 or later:
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm
Start a new session, run /egm:egm-review docs/designs/your-feature.md and pick the "feature spec only" preset to see what fresh reviewers find. The getting started guide also suggests running the method by hand once with three prompts in fresh sessions; Rensin's essay prints them. The plan first, edit never move is the same habit without the plugin. The AGENTS.md, CLAUDE.md and skills post covers where standing rules belong.
Elephant Goldfish model in brief #
- Licence: Apache 2.0, free; it uses your agent's model and plan.
- Install:
agy plugin installfor Antigravity, or a plugin marketplace in Claude Code and Codex. - Runs locally: yes. It is Markdown read by your agent, with no MCP server and no network service of its own.
- Release: v0.1.0 on 27 October 2026, the first tagged release.
New posts by email #
About one email a week with what we published. No account needed.
Which of these do you use? #
Nobody on Stackness runs Elephant Goldfish model yet. Tick what you use and be the first.