Google AI Releases Elephant-Goldfish Model Plugin for Spec-First Coding

Google has open-sourced the Elephant-Goldfish Model plugin, a design-and-review layer that sits in front of AI coding agents. The agent writes a specification first. Fresh subagents with no conversation history then attack that document. Google says thousands of its engineers already use it. The pitch is simple enough to keep: catch a wrong assumption in a Markdown file, not in a pull request.

EGM is not a new language model. It is an Apache 2.0 plugin of skills, reviewer personas, and document standards. The agent supplies the LLM and spends existing quota. It runs on any harness that can spawn fresh subagents, including Antigravity, Claude Code 2.1.271 or later, Claude Cowork, Codex, and ChatGPT. Google published no accuracy benchmark. The evidence on offer is internal adoption and end-to-end harness test runs.

We're watching Google try to fix agent slop at the process layer instead of shipping another coding model. That is the right instinct. Long sessions collapse. Context fills with earlier decisions and generated filler. The Elephant keeps that project history and drafts the spec. It writes no code. Each Goldfish is a stateless reviewer that sees only the draft and the files it references. A same-context review is not a Goldfish review. If a host has no subagent tool, EGM stops.

The five-stage review is the substance. Initial Zoom Out looks at the whole document. A Comprehension Test asks a Goldfish to paraphrase the design — a wrong paraphrase means the doc is ambiguous, which is a sharper check than most human reviews bother with. Critic Review, Intermediate Zoom Out, and Implementation Readiness follow. Each stage gets at most one re-audit. Findings rank as Blocker, Major, or Nit. Upstream, /egm-brainstorm runs a six-lens swarm covering YAGNI, token cost, UX, migration, observability, and tech debt, and writes a concepts brief. /egm-draft turns that brief into a design doc under docs/designs/. Later, /egm-drift-report compares shipped code with the spec; the docs recommend running it weekly.

The published test record looks like plumbing, not a quality score, and that is all we have. Codex 0.163.0 completed 15 native fresh-context reviewers in order, with two live tests passing in 573.13 seconds. Antigravity CLI 1.4.5 ran 16 native reviewers through the stages in order. Codex CLI with the v0.1.0 ZIP passed the full suite in 482.30 seconds. Antigravity CLI, Claude Code, and Codex pass the automated end-to-end fixtures. The Claude Cowork run stopped after stages 5 to 7 on a usage limit. Issue #237 reports AGENTS.md leaking into Codex reviewers. On Claude Code, reviewers get only Read, Grep, and Glob, and CLAUDE.md files are excluded from their context — a sensible hardening that other hosts have not fully matched.

I'm glad this exists. Forcing independent review before code is a real fix for the slop and session context collapse Google's Open Source blog named. It is also more distinctive than Spec Kit or Kiro, which center human review between phases, or Superpowers, which isolates subagents mainly at implementation time. The downside is not theoretical. Token use is heavy. Isolation is unverified on some hosts. The README says this is not an officially supported Google product. Launch coverage described Antigravity-only support, with more coming; the current harness guide now lists Antigravity CLI and App, Claude Code, Claude Cowork, Codex, ChatGPT, omp, and Muse. If this scales without hard isolation guarantees, teams will think they got a Goldfish review and actually get a same-context leak wearing a new name. The useful move is to run the ritual, count the tokens, and check what the reviewer can see.

The method comes from Dave Rensin's essay on elephants and goldfish. The plugin bundles seven skills, nine reviewer personas, six brainstorm lenses, and five review stages — a document standard with a review swarm attached, not a weight file.