Google AI Unveils Elephant-Goldfish Model Plugin for Specification-Driven Coding

Google has unveiled the open-source Elephant-Goldfish Model (EGM) plugin, designed as an architectural framework for AI coding agents.

This innovative plugin mandates that an agent creates a specification prior to embarking on code generation.

Newly instantiated subagents, devoid of prior conversation context, scrutinize the specification meticulously.

Google asserts that it is utilized by thousands of its engineers. The proposition is straightforward: rectify erroneous assumptions within a Markdown document rather than waiting for a pull request.

TL;DR

• Size: Features include 7 skills, 9 reviewer personas, 6 brainstorming lenses, and 5 distinct review stages. There are no model weights; the agent leverages the LLM at its disposal.
  
• Runs on: Compatible with any framework capable of initiating fresh subagents: Antigravity, Claude Code 2.1.271+, Claude Cowork, Codex, ChatGPT. Utilizes your existing model allocation.
  
• Performance: No accuracy metrics have been published; adoption and end-to-end harness test outcomes serve as indications of efficacy.
  
• Best: Codex 0.163.0 successfully executed 15 native fresh-context reviewers in sequence, completing 2 live tests in 573.13 seconds.
  
• Worst: The Claude Cowork instance halted during stages 5 to 7 due to usage limitations. Issue #237 indicates that AGENTS.md is leaking into Codex reviewers.
  
• Bottom line: The standout feature is the enforced independent review prior to coding; the drawback lies in excessive token consumption and uncertain isolation on certain hosts.

What is Google’s Elephant-Goldfish Model plugin?

The EGM is an Apache 2.0-licensed plugin comprising skills, reviewer personas, and documentation standards.

It does not constitute a new language model but rather borrows its methodology from Dave Rensin’s essay on elephants and goldfish.

The Google Open Source blog presents it as a remedy for AI “slop” and the phenomenon of session context collapse. The README clearly states that it is not officially sanctioned as a Google product.

How does the Elephant-Goldfish workflow operate?

The Elephant maintains a lengthy session that encompasses project context and prior decisions; it creates the specification without generating code.

Each Goldfish represents a stateless subagent, which engages only with the draft and its referenced files. You are tasked with resolving trade-offs and endorsing the design.

The core loop consists of 3 commands:

• /egm-brainstorm elucidates the idea and subsequently initiates a 6-lens Goldfish swarm, which evaluates YAGNI, token expenditure, UX, migration, observability, and technical debt. The output is documented in concepts_brief.md.
  
• /egm-draft develops the brief into a design document located underdocs/designs/, section by section.
  
• /egm-review executes the Goldfish review.

/egm-drift-report Later juxtaposes the code with the specification. The documentation advises conducting this comparison weekly.

What occurs during the 5-stage Goldfish review?

The review skill progresses through 5 sequential stages:

1. Initial Zoom Out: Provides a high-level examination of the entire document.
   
2. Comprehension Test: A Goldfish paraphrases the design; if the paraphrase is incorrect, the document is deemed ambiguous.
   
3. Critic Review: Reviewers dedicated to design, structure, and style identify gaps and edge cases.
   
4. Intermediate Zoom Out: Re-evaluates the architecture post-fixes.
   
5. Implementation Readiness: Validates file paths, data structures, and signatures.

Each stage permits a maximum of one re-audit, with findings categorized as Blocker, Major, or Nit. Should a host lack a subagent tool, the EGM halts its progress. A same-context review does not qualify as a Goldfish review.

Source link: Marktechpost.com.