Elephant-Goldfish Model: reviewers with no memory check your AI's design

The Elephant-Goldfish Model, EGM for short, is being shared on social networks, in a short summary like the one we received: no code until the plan survives, no AI grading its own homework, no context collapse halfway through a feature. We read the repository google/elephant-goldfish-model, including the reviewer prompts, the review flow and the FAQ, and compared it with that summary. Some details in the summary differ from the repository.

The Elephant-Goldfish Model (EGM) is an open-source plugin for coding agents under Apache-2.0. A main agent with the full project context, the Elephant, drafts the specification. Fresh subagents without the conversation history, the Goldfish, review it in five stages before any code exists. The human makes the architecture decisions.

Two roles, one rule

The Elephant is the running session with the full project context. It interviews you, proposes options and writes the design document. The Goldfish are fresh subagents. They receive no conversation history, but the document, the files it names and optionally a reading list, and they search the code themselves when needed. Their prompts make them read-only: no writing, no compiling, no fixing.

The reasoning is in the repo's review design. When one context window produces both the plan and the review, the model treats as obvious what exists only in its own memory. Wrong API assumptions from the discussion go untested, and plans without file paths or a test strategy get waved through. The same argument appears in our own working standard, more on that below. How a coding agent works from the task to the test is described in our post Agentic Coding.

The skills

egm-brainstorm tests an idea and writes a concept brief (concepts_brief.md).
egm-draft turns it into design_draft.md and pauses after each section for feedback.
egm-review checks the document with fresh agents.
egm-resolve works through the findings.
egm-bootstrap and egm-drift-report document an existing codebase and compare documents with the current source.

What `egm-review` actually does

The summary we received describes one Goldfish that explains the design back, several hostile reviewers, and a last one that makes the go or no-go call. The review design has five stages and seven personas:

After stages 3 and 5, egm-resolve can work through the findings, depending on the checkpoint. The end point is review by humans. (Quelle: google/elephant-goldfish-model, docs/designs/review.md)

Zoom Out (Skeptical Architect): Is the problem framed correctly? Is anything speculative that should go? Does the codebase already have something to reuse?
Comprehension (Naive Reader): The reviewer paraphrases in its own words what the document proposes. If it cannot paraphrase a section, that is a blocker. According to the prompt, a list of undefined terms will not catch a document whose terms are all defined but whose parts do not connect. This is the "explain the design back" stage from the summary.
Critic: The Design Critic looks for single points of failure, scaling limits and failure modes. Structural checks the document layout, Style checks tone and wording.
Zoom Out, second pass: It flags for removal the defensive bloat that stages 2 and 3 added.
Implementation Readiness: A tech-lead reviewer checks whether it could start coding right away, and a codebase auditor checks paths, classes and function signatures against the real source so the plan contains no invented APIs. Style reads along again.

Stage	Persona	Checks
1	Zoom Out	Problem framing, speculative parts, reuse
2	Comprehension	Paraphrase in own words, comprehension gaps
3	Design Critic, Structural, Style	Failures and limits, layout, tone
4	Zoom Out	Bloat added in stages 2 and 3
5	Implementation Readiness, Codebase Details, Style	Actionability, paths and signatures against the source

Status 2 November 2026, not re-measured by us. Source: review.md and the persona files in the agents/ folder of the repository.

There is no single go or no-go instance. The result is "ready for human peer review".

How findings get resolved

egm-resolve sorts every finding into one of four classes: PROPOSE FIX, CONTEST, ESCALATE and DECLINE for nitpicks. The Elephant may only contest with evidence the reviewer could read itself. Comprehension gaps cannot be contested, and severities do not change without a justified contest. The loop ends after two consecutive clean rounds with no open blocker or major finding, or when the engineer halts it. After five rounds it asks whether to accept, continue or go back to drafting. It also asks when a new finding could only be resolved by reverting an earlier fix. The skill definition of egm-review adds a cap of one re-audit per stage. So the criterion is a count of clean rounds, not zero findings.

Limits the project names itself

No subagents, no review: if the session cannot start fresh subagents, EGM stops, because a review in the same context is not a Goldfish review.
Not an official Google product, as the README says explicitly. It also says thousands of Google engineers use EGM. We cannot verify that, and it is the project's statement about itself.
No proof of time savings: according to the FAQ it depends on the task, and completing the workflow alone does not show a productivity gain. You are told to compare total time, useful findings and rework on your own tasks. The same FAQ warns that too much documentation can become a review bottleneck of its own.
Unsuited to small fixes: the README points to the normal workflow for these.
Data flow: The plugin runs no service and no telemetry. The host account, meaning Claude Code, Codex or Antigravity with their model provider, processes all files and prompts. egm-bootstrap can run local commands and a project formatter, which according to the privacy notice may contact a package registry or another service. Under our own rules that means designs touching WinLine, banking or logistics data go only through local model routes. How we provide local models for company data is described in the service Corporate LLM.

Installation

For Claude Code:

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

Then start a new session. The skills use the prefix /egm:. Codex and the Antigravity CLI have their own commands in the README. According to GitHub the repository was created on 13 September 2026 and had about 10,200 stars when we wrote this.

What we took from it

We evaluated EGM but did not install it. On 29 October 2026 we adopted two rules into our working standard:

A counter-review of a plan, code or our own result always runs in a subagent with an empty context, never in the author's own thread. The reviewer gets the artifact and the task, not our reasoning and not our conclusion.
A review loop stops on convergence, when a new round adds nothing new, rather than waiting for zero findings.

For this we created a read-only reviewer agent that can only read, search and list files. We have no measurements of the benefit yet.

If you try EGM, pick a task with a real design decision, run it once and count afterwards which findings changed the draft. If the document is unchanged after the first round, the effort was not justified for that task. A related approach with an empty chat history in parallel branches is in our post on Fork-and-Flush. More posts on agents and automation in production are collected under the topic AI in production.