Google open-sources Elephant-Goldfish plugin for AI coding specs

The Elephant and the Goldfish

Google has released the Elephant-Goldfish Model (EGM) plugin as an open-source tool for AI coding agents. The project is designed to address 'AI slop' and the problem of session context collapse, where an AI loses track of high-level goals during a long conversation. Instead of jumping straight into writing code, the plugin forces the AI to create a formal specification first.

The workflow is based on a concept by Dave Rensin regarding different types of memory. In this system, the 'Elephant' represents a long-running session that holds the project context and previous decisions. This agent is responsible for drafting the design document but is strictly forbidden from writing any actual code.

The 'Goldfish' agents are the opposite; they are stateless subagents with no access to the conversation history. These agents are tasked with attacking the specification created by the Elephant. By using fresh subagents, the system ensures that the review is unbiased and not influenced by the same assumptions that led to the initial design.

A structured review workflow

The EGM plugin provides a specific set of commands to guide the development process. Users can use a brainstorm command to clarify an idea through six different lenses, such as token cost, user experience, and technical debt. This process results in a concepts brief, which is then turned into a formal design document section by section.

Once a draft exists, the Goldfish review begins. This is a five-stage sequential process that starts with a 'Zoom Out' to check the high-level document. It then moves to a comprehension test where a Goldfish agent must paraphrase the design; if the paraphrase is wrong, the document is flagged as ambiguous.

The final stages include a critic review to hunt for edge cases and an implementation readiness check to verify data structures and file paths. Findings from these stages are ranked as Blockers, Major issues, or Nits. This structured approach ensures that errors are caught in a Markdown file rather than during a complex pull request later in the cycle.

Compatibility and performance

Because EGM is a plugin and not a new model, it does not have its own weights. Users provide their own Large Language Model (LLM) and use their existing quotas. The plugin is designed to run on any harness capable of spawning fresh subagents, including popular tools like Claude Code, Codex, and ChatGPT.

As MarkTechPost reported, early tests show varying results across different platforms. For example, Codex 0.163.0 successfully completed 15 native fresh-context reviewers in order, finishing two live tests in 573.13 seconds. However, other runs, such as those on Claude Cowork, have occasionally stopped due to usage limits during the middle stages of the review.

The plugin is currently supported on several platforms, including Antigravity, Claude Code, and Muse. While it offers a robust way to enforce independent reviews, some users have noted issues with instruction files leaking into the reviewer context on certain command-line interfaces. Google has framed this as a way to ensure that the design is solid before the agent begins the heavy lifting of coding.

Comparison with other tools

EGM distinguishes itself from other design tools like Spec Kit or Kiro by focusing on zero-context agents. While those tools often rely on human intervention between different phases of development, EGM automates the critique process using independent AI subagents. It creates a barrier between the 'thinker' and the 'critic' to ensure the logic is sound.

Another tool, Superpowers, focuses on isolating subagents primarily during the implementation phase. EGM takes a different approach by moving that isolation to the very beginning of the project. By forcing a review of the specification before any code is generated, it aims to save time and resources that would otherwise be wasted on refactoring poorly planned features.

The plugin is intended to be a standard for high-quality agentic workflows. It provides a set of skills, reviewer personas, and document standards that can be applied across different environments. This modularity allows teams to maintain a consistent level of rigor regardless of which specific AI coding tool they choose to use for their daily tasks.