Google has open-sourced the Elephant-Goldfish Model plugin, a design-and-review layer for AI coding agents. The tool forces an agent to write a specification before it writes code. Fresh subagents with no conversation history then attack that spec. Google states thousands of its engineers use it. It is quite simple: catch wrong assumptions in a Markdown file, not in a pull request.
TL;DR
- Size: 7 skills, 9 reviewer personas, 6 brainstorm lenses, 5 review stages. No model weights; your agent supplies the LLM.
- Runs on: any harness that can spawn fresh subagents: Antigravity, Claude Code 2.1.271+, Claude Cowork, Codex, ChatGPT. Uses your existing model quota.
- Performance: no accuracy benchmark published. Evidence is adoption and end-to-end harness test runs.
- Best: Codex 0.163.0 completed 15 native fresh-context reviewers in order. 2 live tests passed in 573.13s.
- Worst: the Claude Cowork run stopped after stages 5 to 7 on a usage limit. Issue #237 reports AGENTS.md leaking into Codex reviewers.
- Bottom line: best thing, forced independent review before code. Worst thing, heavy token use and unverified isolation on some hosts.
What is Google’s Elephant-Goldfish Model plugin
EGM is an Apache 2.0 plugin of skills, reviewer personas and document standards. It is not a new language model. The method comes from Dave Rensin’s essay on elephants and goldfish. The Google Open Source blog frames it as a fix for AI slop and session context collapse. The README states it is not an officially supported Google product.
How does the Elephant-Goldfish workflow work
The Elephant is a long session holding project context and earlier decisions. It drafts the spec and writes no code. Each Goldfish is a stateless subagent. It reads only the draft and the files it references. You resolve trade-offs and approve the design.
The core loop has 3 commands:
- /egm-brainstorm clarifies the idea, then runs a 6-lens Goldfish swarm. Lenses cover YAGNI, token cost, UX, migration, observability and tech debt. Output: concepts_brief.md.
- /egm-draft turns the brief into a design doc under docs/designs/, section by section.
- /egm-review runs the Goldfish review.
/egm-drift-report later compares code with the spec. The docs recommend running it weekly.
What happens in the 5-stage Goldfish review
The review skill runs 5 sequential stages:
- Initial Zoom Out: checks the whole document at a high level.
- Comprehension Test: a Goldfish paraphrases the design. A wrong paraphrase means the doc is ambiguous.
- Critic Review: design, structural and style reviewers hunt gaps and edge cases.
- Intermediate Zoom Out: rechecks architecture after fixes.
- Implementation Readiness: checks file paths, data structures and signatures.
Each stage allows at most 1 re-audit. Findings are ranked Blocker, Major or Nit. If a host has no subagent tool, EGM stops. A same-context review is not a Goldfish review.
Which coding agents does EGM support
Launch coverage described Antigravity-only support, with more coming. The current harness guide covers Antigravity CLI and App, Claude Code, Claude Cowork, Codex, ChatGPT, omp and Muse. Antigravity CLI, Claude Code and Codex pass automated end-to-end fixtures.
Claude Code needs version 2.1.271 or later. Reviewers there get only Read, Grep and Glob tools. CLAUDE.md files are excluded from their context.
The test record gives concrete run data:
- Antigravity CLI 1.4.5: 16 native reviewers ran, stages in order.
- Codex 0.163.0: 15 native reviewers, 2 live tests in 573.13s.
- Codex CLI with the v0.1.0 ZIP: full suite passed in 482.30s.
How does EGM compare with Spec Kit, Kiro and Superpowers
| Feature | Google EGM | GitHub Spec Kit | Kiro Specs | Superpowers |
|---|---|---|---|---|
| Type | Plugin (skills + personas) | Toolkit + Specify CLI | Feature of an agentic IDE/CLI/web | Skills plugin |
| License | Apache 2.0 | MIT | Commercial | MIT |
| Workflow | Brainstorm, draft, review, drift report | Constitution, specify, plan, tasks, implement, converge | Requirements, design, tasks | Brainstorm, plan, subagent build, TDD, review |
| Spec reviewed by fresh-context subagents | Yes, 5 stages before code | Not disclosed | Not disclosed (Analyze Requirements pass) | No; fresh subagents review code per task |
| Harness | Antigravity, Claude Code, Codex, ChatGPT, more | Copilot and other agents | Kiro only | 16 harnesses listed |
| Setup | Plugin install | Python 3.11+, uv | Kiro account | Plugin install |
| Cost | Free plugin; host model usage | Free toolkit; host model usage | Subscription | Free plugin; host model usage |
EGM’s distinct move is reviewing the spec with zero-context agents. Spec Kit and Kiro center on human review between phases. Superpowers isolates subagents mainly at implementation time.
What it means
Developers using these tools will see a shift in how they write code. Instead of letting an AI generate a solution immediately, the system pauses to generate a specification first. This specification then undergoes a rigorous, five-stage review by separate AI agents that have no memory of previous steps. The goal is to catch logical errors and ambiguous requirements before any lines of code are written. The trade-off is increased cost in tokens and time, as the system must spawn multiple fresh instances to perform this independent audit.
