Google's EGM: let the elephant write the spec, let the goldfish tear it apart
The agent that wrote your design doc is the worst possible reviewer of it. It remembers everything you said, so it can't see what the doc forgot to say. Google's Elephant-Goldfish Model, or EGM, is built around that one idea, and it had a loud weekend on X after its v0.1.1 release on Nov 6. The repo sits at about 14,900 stars.
The setup is three roles. The elephant holds your project context and earlier decisions, and drafts the specification. Goldfish reviewers get only the document and the source files it references, with none of the earlier conversation, and their job is to find missing assumptions, unclear requirements and implementation risks. You resolve the trade-offs and approve the spec before any code gets written.
In practice it's a plugin with a handful of skills: brainstorm (fresh reviewers challenge the idea, output goes to concepts_brief.md), draft (the agent writes design_draft.md section by section and waits for your feedback), review (fresh agents check the design across five stages), resolve, bootstrap docs for an existing codebase, and a drift report that compares checked-in docs to current code. It installs into Claude Code, Codex, Antigravity and ChatGPT Work, and it is Apache 2.0. If your agent can't spawn fresh-context subagents, EGM refuses to run the review rather than fake it.
The interesting part is that a Google repo ships first-class for Anthropic's and OpenAI's harnesses. Context isolation is becoming the cheapest quality control there is. You don't need a better model to catch the assumption you never wrote down. You need a model that never heard it.
The README is honest about the limit: a review "does not replace tests or human judgment". Good. The goldfish catches what the document leaves unsaid, not whether the code works.
Source: https://github.com/google/elephant-goldfish-model and the original essay by Dave Rensin: https://medium.com/@drensin/elephants-goldfish-and-the-new-golden-age-of-software-engineering-c33641a48874
← Back to all articles
The setup is three roles. The elephant holds your project context and earlier decisions, and drafts the specification. Goldfish reviewers get only the document and the source files it references, with none of the earlier conversation, and their job is to find missing assumptions, unclear requirements and implementation risks. You resolve the trade-offs and approve the spec before any code gets written.
In practice it's a plugin with a handful of skills: brainstorm (fresh reviewers challenge the idea, output goes to concepts_brief.md), draft (the agent writes design_draft.md section by section and waits for your feedback), review (fresh agents check the design across five stages), resolve, bootstrap docs for an existing codebase, and a drift report that compares checked-in docs to current code. It installs into Claude Code, Codex, Antigravity and ChatGPT Work, and it is Apache 2.0. If your agent can't spawn fresh-context subagents, EGM refuses to run the review rather than fake it.
The interesting part is that a Google repo ships first-class for Anthropic's and OpenAI's harnesses. Context isolation is becoming the cheapest quality control there is. You don't need a better model to catch the assumption you never wrote down. You need a model that never heard it.
The README is honest about the limit: a review "does not replace tests or human judgment". Good. The goldfish catches what the document leaves unsaid, not whether the code works.
Source: https://github.com/google/elephant-goldfish-model and the original essay by Dave Rensin: https://medium.com/@drensin/elephants-goldfish-and-the-new-golden-age-of-software-engineering-c33641a48874
Comments