Google's Elephant-Goldfish Plugin: Spec-First AI Coding, Reviewed by Agents With No Memory

AI coding agents have a speed problem that looks like a quality problem. They write code fast, and when the plan in their head is wrong, they write the wrong code fast too, then spend the rest of the session defending it. Google's answer, open-sourced on October 22, 2026, is a plugin that forces the agent to slow down at exactly one point: before any code exists. The Elephant-Goldfish Model, or EGM, makes the agent draft a design document first, then hands that document to fresh subagents that have never seen the conversation and asks them to break it.

It is not a model. There are no weights to download and no API to call. EGM is a set of skills, reviewer personas and document standards that plug into the coding agent you already use. The repository passed 14,579 stars and 1,070 forks within about seven weeks of its September 13, 2026 creation, with version 0.1.1 shipping on November 6. That is attention, not proof, and this article keeps the two separate.

What the Elephant-Goldfish Model actually is

The idea comes from an essay by Dave Rensin, one of the plugin's authors, called "Elephants, Goldfish, and the New Golden Age of Software Engineering." Google's Open Source blog, which announced the plugin, puts the problem in three words: slop, bugs and context collapse. An agent that has been in a long session shares your blind spots. It knows what you meant, so it stops noticing what you actually wrote.

EGM splits the work between two kinds of session. The Elephant is a long, context-rich session that holds the project, the requirements and every earlier decision. It interviews you, debates trade-offs and drafts the specification, and it writes no code. The Goldfish are stateless subagents that see only the draft and the files it references. Because they share none of the designer's context, they act as objective critics, probing for gaps and edge cases. You sit between them: you resolve the trade-offs and approve the specification before anything is implemented.

Google says the plugin was developed internally for its own engineers and that thousands of Googlers use it every day to ship production code. That adoption claim is Google's own. The company has published no figures for defects caught or rework avoided, and no independent evaluation of the method exists yet. Treat the internal-use story as a reason to try it, not as evidence it works.

How the workflow works

The plugin automates a three-phase lifecycle, driven by slash commands. Launch coverage counts seven skills, nine reviewer personas, six brainstorm lenses and five review stages in the package.

Phase 1: brainstorm

/egm-brainstorm opens the conversation with a "No Code" rule. You describe the feature, the agent clarifies the idea, and a swarm of Goldfish reviewers examines it through six lenses: YAGNI, token cost, user experience, migration, observability and technical debt. The agreed direction is saved as concepts_brief.md.

Phase 2: draft

/egm-draft turns the brief into a design document under docs/designs/, section by section, pausing after each section for your feedback. The document covers the problem definition, the technical plan, the alternatives considered and a detailed implementation plan that names the specific files to change.

Phase 3: review

/egm-review runs the Goldfish protocol on the draft. The review skill runs five sequential stages: an initial zoom out over the whole document, a comprehension test in which a Goldfish paraphrases the design back (a wrong paraphrase means the document is ambiguous), a critic review by design, structural and style reviewers, a second zoom out after fixes, and an implementation-readiness check that verifies file paths, data structures and function signatures. Findings are ranked Blocker, Major or Nit, and each stage allows at most one re-audit. A separate /egm-drift-report command compares checked-in documents with the current source code. The docs recommend running it weekly.

Two rules matter. First, if your agent cannot create fresh subagents, EGM stops the review: a same-context review is not a Goldfish review. Second, a review does not replace tests or human judgment. The README says so directly.

How to install it

Install the complete plugin so its skills can read the reviewer prompts and document standards. The commands below come from the project's README, read in full on November 8, 2026.

For Claude Code (version 2.1.271 or later), run claude plugin marketplace add google/elephant-goldfish-model followed by claude plugin install egm@egm. For Codex CLI, the pair is codex plugin marketplace add google/elephant-goldfish-model and codex plugin add egm@egm. For the Antigravity CLI, run agy plugin install https://github.com/google/elephant-goldfish-model. ChatGPT and Codex desktop users install it from the published plugin listing in the desktop app; that listing is desktop-only. Start a new session after installation. In Antigravity and Claude Code, prefix the commands with /egm:, for example /egm:egm-brainstorm. In Codex, pick the matching skill from the skill picker.

Support beyond the first three harnesses is wider than the launch coverage suggested. The current harness guide covers the Antigravity CLI and app, Claude Code, Claude Cowork, Codex, ChatGPT, omp and Muse. The project's own test records report that Antigravity CLI 1.4.5 ran 16 native reviewers through the stages in order, and Codex 0.163.0 completed 15 native fresh-context reviewers with two live tests in 573.13 seconds. A Codex CLI run with the v0.1.0 ZIP passed the full suite in 482.30 seconds. Those are the project's own execution records, not independent benchmarks. They show the plumbing works, not that the reviews catch real bugs.

There are also extra skills for the rest of the lifecycle: /egm-resolve works through review findings, /egm-bootstrap creates or updates documentation for an existing codebase, and /egm-review-loop runs a configured group of reviewers for a specific review task. You can use any skill on its own; the full sequence is optional.

What it costs

The plugin itself costs nothing. It is released under the Apache 2.0 license, and the repository states plainly that it is not an officially supported Google product, so there is no support contract to buy and no hosted version to subscribe to. The real cost is the model usage behind it, because every Goldfish reviewer is a fresh subagent call on your existing coding-agent plan.

Claude's subscription and API prices above were verified on claude.com/pricing on November 8, 2026. Sterling and Canadian dollar figures use November 8, 2026 market rates ($1 = £0.747 / CA$1.41) and are approximations; check the current rate before budgeting. If you run EGM in Codex or ChatGPT instead, the cost lands on your existing OpenAI plan. The plugin adds no charge of its own. Either way, the review phase is the expensive part of the workflow: five stages of fresh-context reviewers, each reading the full draft, consume noticeably more tokens than a single drafting pass. That is the trade the plugin asks you to make, more tokens now for fewer wrong assumptions baked into code later.

How it compares with Spec Kit, Kiro and Superpowers

EGM is not the first spec-first workflow for coding agents. What sets it apart is where the review happens, before code, by agents with no memory of the design conversation. The comparison below draws on launch coverage of all four tools.

Spec Kit and Kiro center on human review between phases; Superpowers isolates subagents mainly at implementation time. EGM's distinct move is isolating the reviewers at the design stage, when a caught mistake is still a paragraph rather than a pull request. For a fuller look at the agent it most often runs on, see our Claude Code explainer, and for the open-source alternative to Claude Code itself, see our OpenCode review. If you are choosing between coding agents first, our roundup of the 7 best AI coding agents in 2026 compares prices and limits side by side.

The honest catches

EGM has real limits, and most of them are visible in the project's own records rather than hidden.

There is no accuracy benchmark. Google has published no study showing that Goldfish reviews catch more defects than an ordinary prompt, or that specs written with EGM produce better code. The evidence on offer is adoption, thousands of internal users, plus end-to-end harness test runs. That is a weak form of evidence, and the plugin's own documentation does not pretend otherwise.

Reviewer isolation leaks on some hosts. Issue #237 in the project tracker reports AGENTS.md leaking into Codex reviewers, which breaks the zero-context promise on that host. The Claude Cowork test run stopped after stages 5 to 7 on a usage limit, so the full five-stage review does not always complete there. And on Claude Code, reviewers get only Read, Grep and Glob tools, with CLAUDE.md files excluded from their context. That is a deliberate restriction, but one that limits what they can check.

It is also Antigravity-first by design. The announcement says the plugin's goal is to bring EGM into Antigravity, with other environments upcoming, so the smoothest experience is on Google's own agent platform, and support elsewhere arrived after the fact.

Finally, remember what it is not: not a model, not a Google product with support, and not a substitute for tests. A spec that survives five Goldfish can still be wrong in ways only running code reveals.

Who should try it, and who should skip it

EGM fits work where the design decision is the expensive part: a new service, a data-model change, an API contract, a migration with rollback risk. If a wrong choice costs a week of rework, spending extra tokens on five reviewers is cheap insurance. Teams already living in Claude Code or Codex get the most from it, since installation is two commands and the review loop needs no new account.

Skip it for one-line fixes and trivial changes. The README itself tells you to. Skip it, too, if your agent cannot spawn fresh subagents: without that, the review cannot run at all, and a same-context pass adds cost without the benefit. And if your budget is token-sensitive, run it on the features that matter and leave the rest alone. A full five-stage review on every change would quietly become your largest API line item.

Common mistakes

The first mistake is treating the review as testing. A Goldfish can tell you the spec is ambiguous; it cannot tell you the code works. Run the tests anyway. The second is reviewing in the same context as the draft. If the reviewer shares the Elephant's memory, you have built a rubber stamp, not a critic. The third is skipping the human step: the workflow ends with you approving the design, and approving without reading defeats the purpose. The fourth is installing it and never running the drift report. Specs rot, and a weekly /egm-drift-report is the only part of the plugin that notices.

Is the Elephant-Goldfish Model an AI model?

No. It is an open-source plugin of skills, reviewer personas and document standards. It has no weights and no API of its own; your coding agent supplies the language model, and you pay that agent's normal usage rates.

Does EGM work with Claude Code?

Yes, with Claude Code version 2.1.271 or later. Install it with claude plugin marketplace add google/elephant-goldfish-model and claude plugin install egm@egm, then start a new session. It also runs on Codex, ChatGPT, Claude Cowork and Antigravity.

How much does the Elephant-Goldfish plugin cost?

The plugin is free under the Apache 2.0 license. Your only cost is the model usage of the host agent — for example, Claude Pro at $20 a month (about £14.94 / CA$28.20) or Sonnet 5.5 API usage at $2 input and $10 output per million tokens, verified November 8, 2026.

Is EGM an officially supported Google product?

No. The repository states that it is not an officially supported Google product. It was developed for Google's internal engineers, open-sourced under Apache 2.0, and is maintained as a community open-source project.

Does EGM replace code review or tests?

No. It reviews the design document before code is written; it does not review code and does not run tests. The project documentation states directly that a review does not replace tests or human judgment.

Google's Elephant-Goldfish plugin is a small idea with a large implication: the most valuable thing an AI coding agent can do before writing code is argue with the plan. The plugin is free, the install takes minutes, and the review loop is different from anything Spec Kit, Kiro or Superpowers do at the design stage. The catch is that nothing yet proves the argument makes the code better, only that Google's engineers use it and the plumbing works. Try it on one real feature, read what the Goldfish find, and judge it by whether the spec that comes out is one you would have written yourself.

Sources

• Google Open Source Blog — Introducing the Elephant-Goldfish Model plugin (October 22, 2026)
• GitHub — google/elephant-goldfish-model (README, install commands, skills list; stars and releases verified via GitHub API, November 8, 2026)
• egm.dev — Elephant-Goldfish Model project site
• MarkTechPost — Google AI Releases Elephant-Goldfish Model Plugin for Spec-First Coding (November 8, 2026)
• Claude — Pricing (Pro $20/month, Max from $100/month, Sonnet 5.5 API $2/$10 per million tokens; verified November 8, 2026)