Google's Elephant-Goldfish Model (EGM) for AI Coding, Explained

AI coding agents write code fast. Google’s engineers say that speed has a cost: according to the team’s announcement, it “has silently introduced a potential reliability crisis by also increasing the velocity of ‘slop,’ bugs, and session context collapse.”

Their answer is the Elephant-Goldfish Model (EGM), released as open source on 22 October 2026. Here’s what it is, how it works and how to try it.

The idea in one sentence

Design with an elephant, review with a goldfish.

• The elephant never forgets. It is one long-running AI session that holds the whole project in mind (the architecture, your intent and the small decisions made along the way) and turns them into a written design. ==It writes no code.==
• A goldfish remembers nothing. Each goldfish is a brand-new, stateless session that reads only the design document and then checks it.

You ──► 🐘 Elephant (long memory)
          brainstorm → draft design doc
                     │
                     ▼
        🐟 🐟 🐟  Goldfish reviewers (no memory)
          understand it? · attack it · ready to build?
                     │
          problems ──┘ back to the Elephant to fix
                     │
                     ▼ approved design
              Coding agent writes the code

Why a reviewer with no memory?

When the same AI session that designed something also reviews it, it carries the same assumptions and blind spots, so it tends to agree with itself. A goldfish has seen none of the earlier conversation. If the design only makes sense to someone who was there, the goldfish will notice.

It’s the AI version of an old engineering habit: ask someone who wasn’t in the meeting to read the spec. A design that a fresh reader can understand and implement is a design that isn’t relying on hidden context.

How the workflow runs

The plugin gives your coding agent a set of commands (Google calls them skills). The three core phases are:

1. Brainstorm with the elephant (egm-brainstorm)

You and the elephant talk through the problem under a no-code rule: debate options, question assumptions and settle decisions. The agreed direction is saved as a concepts brief (concepts_brief.md).

2. Draft the design (egm-draft)

The elephant writes a Markdown design document section by section, pausing for your feedback. Per Google’s description, it covers the problem, the technical plan, the alternatives considered and an implementation plan that names the specific files to change.

3. The goldfish review (egm-review)

Fresh goldfish sessions review the design. Google’s announcement describes three checks:

1. Comprehension check: a goldfish reads the document and explains the design back. If it gets the plan wrong, the document is ambiguous and goes back for more work.
2. Antagonistic review: several goldfish act as hostile reviewers, hunting for edge cases, hidden dependencies and logical gaps. Their objections go back to the elephant to resolve.
3. Implementation readiness: a final goldfish makes a go/no-go call on whether file paths, data structures and function signatures are clear enough to build from.

Only after the design passes does a coding agent implement it. Approved designs are kept in the repository (under docs/designs/), so the reasoning stays with the code.

The repository also includes helper skills, such as working through review findings (egm-resolve), documenting an existing codebase (egm-bootstrap) and comparing documents against the current code to spot drift (egm-drift-report).

How to install it

The project lives at github.com/google/elephant-goldfish-model, with docs at egm.dev. At release, the README lists:

# Antigravity CLI
agy plugin install https://github.com/google/elephant-goldfish-model

# Claude Code
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

# Codex
codex plugin marketplace add google/elephant-goldfish-model
codex plugin add egm@egm

Start a new session afterwards and run a skill, for example /egm:egm-brainstorm. One requirement: brainstorming and review need an agent that can create fresh sub-agents, because that’s how the goldfish are made. Check the README for the latest commands, as they may change.

When to use it (and when not to)

Good fit:

• New features, services or refactors that touch several files
• Work an AI agent will implement mostly on its own
• Teams that want a reviewable record of why a design was chosen

Skip it for: small, obvious bug fixes. The README itself suggests using your normal workflow for those.

Limitations to know

• It isn’t a code review. It checks the design before coding. You still need tests and normal review of the code the agent writes. The README says plainly: “A review does not replace tests or human judgment.”
• The workflow stops at the handoff. It makes sure the design is complete and clear, but it doesn’t verify that the coding agent followed it faithfully.
• Security isn’t a named review step. Coverage notes that security isn’t on the default reviewers’ list by name, though the reviewer prompts can be customised to add one.
• No published metrics yet. Google says thousands of its engineers use EGM every day, but hasn’t published before-and-after numbers on bugs or review time.
• Not an officially supported Google product. It’s Apache 2.0 open source, copyrighted by Google, without a formal support commitment.

The bigger lesson for AI-assisted coding

You don’t need the plugin to borrow the idea:

1. Separate thinking from typing. Agree the design before letting an agent write code.
2. Write the design down in a document that could stand on its own.
3. Review with fresh eyes. Open a new AI session (or ask a colleague) with only the document and ask it to explain the plan back, then to attack it.
4. Hand the approved document to the coding agent, and still review the code.

It’s the same principle as good RAG and agent design: context is powerful, but too much shared context hides mistakes. A reviewer with no memory is a cheap way to find them.

વધુ જાણો

• Google’s announcement: Introducing the Elephant-Goldfish Model plugin
• Code and install guide: google/elephant-goldfish-model on GitHub
• Our guides: How to build an AI agent, AI coding assistants and agents and the AI agents roadmap