Google open-sources Elephant-Goldfish plugin for AI coding

The Elephant and the Goldfish

Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, a new design-and-review layer for AI coding agents. The project is based on a concept by Dave Rensin regarding how different types of memory affect information processing. In this workflow, the 'Elephant' represents a long session that holds the project context and previous decisions.

The Elephant is responsible for drafting the project specification but is strictly prohibited from writing any actual code. This separation ensures that the high-level goals of the project remain stable and are not diluted by the technical details of implementation. By focusing purely on the plan, the Elephant provides a clear roadmap for the next steps.

In contrast, the 'Goldfish' represents a series of stateless subagents that have no memory of previous conversations. These subagents are tasked with attacking the drafted specification to find flaws. Because they lack context, they are less likely to be influenced by the 'slop' or biases that can accumulate in long AI chat histories.

A spec-first coding workflow

The EGM plugin introduces a specific sequence of commands to guide the development process. Developers can use the /egm-brainstorm command to clarify an idea through a swarm of Goldfish reviewers. These reviewers look at the project through six different lenses, including user experience, token costs, and technical debt, to produce a concepts_brief.md file.

Once the brief is established, the /egm-draft command turns that information into a formal design document. This document is organized section by section within a dedicated directory. This step ensures that the requirements are fully fleshed out and agreed upon before the AI begins the actual construction of the software.

The final core loop involves the /egm-review command, which initiates a rigorous five-stage review. These stages include an initial zoom out to check the high level, a comprehension test where a Goldfish paraphrases the design, and a critic review for edge cases. The process concludes with an implementation readiness check to verify data structures and file paths.

Compatibility and performance

EGM is an Apache 2.0 plugin, meaning it is not a new model and does not include any model weights. Instead, it functions as a layer of skills and personas that runs on top of your existing tools. It is designed to work with any harness capable of spawning fresh subagents, such as Antigravity, Claude Code, Codex, and ChatGPT.

As marktechpost.com reported, the plugin has already seen adoption by thousands of Google engineers. Testing data shows that Codex 0.163.0 completed 15 native fresh-context reviewers in a single run. Other tests show the Antigravity CLI successfully running through all review stages in 573.13s.

While the plugin is highly versatile, it does have some technical requirements and limitations. It requires Python 3.11+ and the uv tool for installation. Additionally, some users have noted that the plugin can be heavy on token usage. There are also reports of some command-line interfaces leaking host instruction files into the reviewer context.

How it differs from competitors

EGM occupies a unique space compared to other agentic tools like Spec Kit, Kiro, and Superpowers. While many existing tools focus on human review between phases, EGM emphasizes automated, independent review by zero-context subagents. This move is intended to solve the problem of 'session context collapse,' where an AI becomes confused by its own previous outputs.

Unlike Superpowers, which primarily isolates subagents during the implementation phase, EGM isolates them during the planning phase. This ensures that the blueprint of the code is sound before a single line is written. By catching wrong assumptions in a Markdown file rather than a pull request, it aims to save significant time during the development cycle.

The plugin is currently free to use, though users must pay for their own model usage. It provides a structured way to handle complex migrations and observability requirements. By providing a standardized set of reviewer personas and document standards, it creates a consistent quality gate for any AI-assisted coding project.

What comes next for developers

For developers looking to adopt EGM, the first step is setting up a compatible harness. The plugin is currently supported on Antigravity, Claude Code (version 2.1.271 or later), Codex, and ChatGPT. Users can find the documentation on the Google Open Source blog to begin integrating the /egm-brainstorm and /egm-draft commands into their workflows.

The project is still evolving, with more harnesses expected to be added in the future. Developers are encouraged to run the /egm-drift-report command weekly to compare the final code against the original specification. This helps ensure that the AI does not deviate from the agreed-upon design as the project grows in complexity.

Because the plugin relies on the host's model, the performance and accuracy will depend on the underlying LLM chosen by the user. However, the structured review process provides a safety net that can help mitigate the hallucinations common in large-scale coding tasks. It represents a shift toward more deliberate, specification-heavy AI engineering.

Key takeaways

The Elephant-Goldfish Model (EGM) is a plugin designed to improve AI coding by enforcing a 'spec-first' approach. It uses an 'Elephant' agent to draft a project plan and 'Goldfish' subagents to critique that plan without the bias of previous conversation history.

The workflow includes five sequential review stages, including a comprehension test and an implementation readiness check. This method is designed to catch logic errors and architectural gaps in a Markdown file before any code is actually generated. It is compatible with major tools like Claude Code and Codex.

While it requires more tokens and careful management of host instructions, EGM offers a way to prevent AI 'slop.' By forcing independent reviews of the design, it helps developers maintain a clear project vision over long and complex development cycles.

Try it

If you use Claude Code or Codex, check the EGM GitHub repository to see if your current version meets the requirements for the /egm-review command.

Watch

How other major AI coding tools respond to the 'spec-first' movement.

Whether the token costs of the 5-stage review become prohibitive for small projects.

Updates on the reported instruction leaks in certain CLI environments.