Google's Elephant-Goldfish Model Reviews Specs With Fresh Agents

Most AI-written designs fail the same way: the agent that wrote the plan is also the one grading it. Google's new open-source plugin, the Elephant-Goldfish Model (EGM), is built around breaking that loop. On Nov. 7, 2026, the Google Cloud Tech account announced it on X, describing a workflow where deep-context agents draft specs and zero-context subagents stress-test them before any code gets written, with the engineer approving the spec.

The code is in google/elephant-goldfish-model under Apache 2.0, at version 0.1.1. The tweet's one-liner is accurate but thin, so here is what the repository shows underneath it.

Google Cloud Tech's Nov. 7 announcement of the Elephant-Goldfish Model plugin (@GoogleCloudTech)

View on X

Three roles, one rule: nobody grades their own work

The Elephant is your normal long session. It knows the codebase, the earlier decisions and your constraints, and it writes the design document. The Goldfish is a fresh subagent that gets only the draft and the files it points to. You are the third role, and the README gives you the call on architectural trade-offs and the final approval.

The pitch only works if the Goldfish really starts blank. EGM treats that as a hard requirement: if your agent can't spawn fresh subagents, the review skill stops and says a review in the same context isn't a Goldfish review, per the review skill's instructions.

Seven skills, and the review has five stages

The plugin ships seven commands. /egm-brainstorm interviews you, then sends a seven-lens swarm of reviewers at your idea and writes a concepts_brief.md. /egm-draft turns that brief into a design document under docs/designs/, one section at a time, pausing for your feedback. /egm-review runs the draft through five phases: a zoom-out pass, a comprehension check, a design critique, a second zoom-out and an implementation-readiness check. The architecture document names the reviewer personas, from a Design Critic and a Codebase Auditor to a Style Editor and a Synthesizer that merges the reports.

Three more skills cover the edges: /egm-resolve walks you through the findings, /egm-bootstrap generates README files for an existing codebase, and /egm-drift-report compares your checked-in specs with the current code. The getting-started guide suggests running the drift report weekly and warns that the review produces "a lot" of feedback on purpose, so you are expected to ignore some of it.

Where it runs

On Claude Code the install is two commands: claude plugin marketplace add google/elephant-goldfish-model, then claude plugin install egm@egm. Codex CLI mirrors that with codex plugin, and Antigravity uses agy plugin install with the repo URL. The harness page also lists a desktop ChatGPT and Codex plugin, Claude Cowork, and community paths for omp and Muse.

The tweet says it "works with" four tools, and the support table is narrower. Standalone Antigravity IDE is listed as not supported because it doesn't expose the subagent tool the reviews need. Claude Cowork and ChatGPT Work have working install paths but open gaps in cloud trace verification. Only Antigravity CLI, Claude Code and Codex are described as passing the project's automated install and end-to-end fixtures.

The "blank slate" has a leak

Here is the part the announcement skips. The same harness page says reviewer isolation evidence is "narrower" than the install results. On Claude Code, the reviewer adapters set omitClaudeMd: true, so a Goldfish doesn't see your project's CLAUDE.md. On Codex and Antigravity, issue #237, opened Oct. 30, reports that reviewers still receive the project and user AGENTS.md files (and GEMINI.md on Antigravity) even though they never see the conversation. The repo's own support notes call these reported host-instruction leaks, not conversation leaks, and say no exclusion control has been verified for either host.

Why that matters: those files often hold a team's standing assumptions, such as what's out of scope. A reviewer that reads them inherits the author's priors, which is the bias the method exists to remove. We read this as a sound design with a host-dependent guarantee: strongest on Claude Code, weaker on the other two.

There's a second nuance. The project site says Goldfish agents read the draft spec and the live source code. So they are blank on your conversation, not on your repository, and a design that misdescribes the code can be caught for exactly that reason.

Scale, cost and who is behind it

GitHub's API showed about 13,900 stars, roughly 1,000 forks and 79 open issues on the day of the announcement, for a repository it lists as created Sept. 13, 2026. Stars measure curiosity, not use. The site says "thousands of Google engineers" use the plugin, a claim we couldn't verify independently. The README also says it is not an officially supported Google product, and the contribution rules require a named human to review every pull request and ban autonomous PRs.

On cost, there's no pricing, since the plugin is free, but there is a clue about token spend. The harness page records a Muse test where one uninterrupted review launched 15 fresh-context reviewers. Each one is a separate model session, so we'd expect a full five-stage review to cost several times a single chat pass. That estimate is ours; the repository publishes no token figures.

Try it on one design, not the whole backlog

If you use Claude Code, pick one change that needs real decisions, like retries and duplicate events in a webhook feature, which is the README's own example. Run the three core commands and compare what the Goldfish found with what you'd have caught yourself. If you run Codex, the caveat above applies: should your AGENTS.md carry strong opinions, the review isn't fully independent. Teams already handing tasks between agents, as in DeepSeek Harness 0.2.2's hand-offs to Codex and Claude Code, will also want to check which instruction files each agent injects.

The project cites Dave Rensin's Medium essay as its basis. We haven't been able to read it, so this account rests on the repository and the project site. The date to watch is the status of issue #237: if the Codex and Antigravity leak gets a verified fix, the strongest objection to this review method goes away.

Source: Google Cloud Tech on X

google elephant goldfish model coding agents spec driven development claude code

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.