WORLDTECHGoogle open-sources a plugin that forces coding agents to write specs first
Google has made its Elephant-Goldfish Model (EGM) plugin available as open source, a layer that has AI coding agents draft and interrogate a written specification before any code appears. According to MarkTechPost, the goal is to surface faulty assumptions in a Markdown file rather than in a pull request.
How the plugin operates
Under the plugin, a coding agent produces a specification ahead of any implementation. Fresh subagents that hold no earlier conversation then probe that draft, so the review is not colored by the context that generated it. The design rests on an Elephant, a long-running session that keeps project context and earlier decisions, and on Goldfish reviewers that see only the document placed before them.
The Elephant writes the specification but no code, reading only the draft and the files it points to. People settle trade-offs and sign off on the design.
Commands and review stages
The core loop runs through three commands. A brainstorming command refines an idea and then launches a Goldfish swarm of six lenses: YAGNI, token cost, UX, migration, observability and technical debt. A drafting command builds a design document under docs/designs/ one section at a time, and a later drift-report command checks the code against the specification.
The review skill passes through five stages in order: an initial overview of the entire document, a comprehension test in which a Goldfish restates the design, a critical review by design, structural and style reviewers, an intermediate overview that re-examines the architecture after fixes, and a readiness check of file paths, data structures and signatures. Each stage allows at most one re-audit.
Compatibility and known problems
The plugin works on any environment able to start fresh subagents. The source names Antigravity, Claude Code from version 2.1.271, Claude Cowork, Codex and ChatGPT. It ships no model weights, so the language model comes from the user's own agent.
The source records two problems. One end-to-end run on Claude Cowork stopped after stages five to seven because of a usage limit, and issue #237 reports AGENTS.md leaking into the Codex reviewers. The source also warns of heavy token consumption and isolation that has not been verified on some hosts.
Origin and status
The method is drawn from an essay on elephants and goldfish by Dave Rensin. The Google Open Source blog presents it as a remedy for low-quality AI output, which it calls slop, and for the decay of session context. The README states that EGM is not an officially supported Google product. According to the source, Google says thousands of its engineers use it, and the evidence offered is adoption and end-to-end harness test runs.
Open questions
- Will the AGENTS.md leak into Codex reviewers reported in issue #237 be fixed?
- How well are the subagent reviews isolated on hosts where isolation has not been verified?