What was released
Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, described as a design and review layer intended for agents that write code. According to the source, the goal is to invert the order of work: the agent first produces a specification and only then the code. The plugin is distributed under the Apache 2.0 license and bundles skills, reviewer profiles, and documentation standards; it is not a new language model and does not include weights: the model is provided by the agent hosting it. The README, the source reports, states that it is not an officially supported Google product.
The method: one elephant and many goldfish
The structure draws on an essay by Dave Rensin. The elephant is the long-running session that preserves the project context and decisions already made: it drafts the specification without writing code. The goldfish are subagents with no conversation history, which read only the draft and the files it references. Humans remain in the loop to resolve trade-offs and approve the design, with three main commands dedicated to brainstorming, drafting, and review. An additional command later compares the code with the specification, and the documents recommend running it weekly.
The review is structured in five sequential phases, with at most one re-check per phase and findings classified as blocking, major, or minor. If the environment lacks a tool for generating subagents, the plugin stops: a review in the same context is not considered valid.
Claimed adoption and limited evidence
Google says thousands of its engineers use the system, and the source cites adoption and end-to-end tests as the main evidence. No accuracy benchmark appears to have been published. In the cited tests, Codex version 0.163.0 completed fifteen native fresh-context reviewers in sequence, with two live tests passed in 573.13 seconds; the run on Claude Cowork, by contrast, was interrupted between the fifth and seventh phases because of a usage limit. The source also flags issue number 237, concerning AGENTS.md files that end up in the context of Codex reviewers.
The platform picture has changed compared with the initial presentation, which mentioned support for Antigravity only: the current guide lists Antigravity CLI and App, Claude Code, Claude Cowork, Codex, ChatGPT, omp, and Muse, while Antigravity CLI, Claude Code, and Codex pass the automated tests provided. On Claude Code, from version 2.1.271, reviewers have access only to the Read, Grep, and Glob tools, and CLAUDE.md files are excluded from their context.
Compared with approaches such as Spec Kit and Kiro, which place human review between phases, and Superpowers, which isolates subagents especially during implementation, the claimed difference is specification review entrusted to zero-context agents. The actual isolation on some hosts and the token cost, which the source describes as high, remain to be verified. The plugin consumes the credit of the model already in use.
