Google open-sources tool to catch AI coding errors before they happen

The problem of AI slop

Google is releasing a tool called the Elephant-Goldfish Model (EGM) to address a growing issue in software engineering. While generative AI can significantly increase the speed at which developers write code, it also accelerates the production of 'slop.' This term refers to the bugs, errors, and lack of structure that occur when AI generates large amounts of code without proper oversight.

The team behind the plugin argues that pure speed without structured guardrails creates a reliability crisis. They believe that current AI tools are often just scaling mistakes rather than simply making the work faster. To fix this, they want to align the human need for understanding a system with the AI's ability to execute tasks.

The goal of the EGM plugin is to create a paradigm shift in how software is built. Instead of letting an AI agent jump straight into writing code, the tool forces the AI to think through the design first. This ensures that the final product is based on a solid plan rather than a series of disconnected, automated decisions.

Elephants and Goldfish

The system uses a unique naming convention to describe two different types of AI roles. The 'Elephant' is a long-running session that holds the entire project in view, including the architecture, the intent, and the small decisions made during the process. The Elephant is specifically designed to write no code; its only job is to interview the developer and argue over choices.

In contrast, the 'Goldfish' agents are stateless, which means they start every interaction with no memory of previous conversations. Google's logic is that a reviewer who did not participate in the initial design conversation is less likely to carry over the same blind spots. By using fresh AI sessions, the system ensures a more objective critique of the proposed plan.

This distinction creates a dynamic where the Elephant acts as the project's memory and advocate. It keeps the record of why certain paths were taken while the Goldfish act as independent auditors. This separation of concerns helps prevent the AI from simply agreeing with its own previous logic, which is a common pitfall in long AI conversations.

A three-phase workflow

The plugin operates through three distinct commands that a developer must trigger. The first phase, egm-brainstorm, allows the developer and the Elephant to debate options under a strict no-code rule. During this stage, they test assumptions and explore different technical paths without actually building anything.

The second phase, egm-draft, requires the creation of a Markdown design document. This document must cover the problem, the technical plan, the alternatives considered, and a specific implementation plan that names the exact files to be changed. The AI cannot move forward until this comprehensive blueprint is established and agreed upon.

The final phase, egm-review, brings in the Goldfish agents to stress-test the draft. One Goldfish must explain the design back to the Elephant to check for clarity. If it fails to explain the plan correctly, the document is flagged as ambiguous and sent back for more drafting. Other Goldfish then hunt for edge cases and logical gaps before a final 'go/no-go' call is made.

Customizable guardrails

While the plugin is designed to catch general logic errors and hidden dependencies, it is highly customizable. The repository includes Markdown rules and defined personas for the different subagents. This means a team can modify the roles the Goldfish play to suit their specific needs for a project.

For example, if a team requires a specific threat review, they can set up a security-minded Goldfish persona. However, security flaws are not automatically included in the standard review list; the team must intentionally configure the plugin to prioritize them. This gives developers control over what specific types of errors they want the AI to hunt for during the review phase.

The tool currently works inside the Antigravity development environment, though Google plans to support other environments in the future. By making the plugin open-source, Google is providing a framework that other companies can adapt to their own internal development workflows to manage AI-driven code sprawl.