Google’s example starts with a feature that sounds simple: add webhook delivery to a software project. A webhook sends another service a message when something happens. But what should the software do if delivery fails, or the same event arrives twice? Those decisions matter before an AI agent starts changing files.
The Elephant-Goldfish Model, or EGM, is Google’s open-source plugin for working through those questions. One AI session helps develop the plan with the project’s history in view. Fresh reviewers then challenge the written plan without inheriting the conversation that produced it. A person decides which criticism matters and approves the design before implementation.
Despite its name, EGM is not a new AI model competing with Gemini, Claude or GPT. It is a way of organising existing assistants: instructions, reviewer roles and document standards packaged into a plugin. The aim is to catch a bad decision before faster code generation makes it expensive to undo.
Google Cloud Tech’s demonstration promotes a project Google open-sourced in October.
The elephant remembers the discussion. The goldfish has to read the plan.
An AI assistant’s context is the information it can use while answering: messages, instructions and files it has been given. A coding agent adds tools that can inspect a project, edit files and run commands. A subagent is a separate assistant session assigned part of the work.
EGM’s “elephant” holds the requirements, project history and design discussion. It asks questions and helps write a specification, meaning a document explaining what the software should do and how the team intends to build it. The “goldfish” reviewers get the document and relevant source material, but not the earlier back-and-forth.
A decision agreed in chat can still be missing from the specification. A reviewer who already heard it may fill in the gap. A fresh reviewer has to work from what is actually written. For another engineer, or another agent taking over later, those gaps are real.
In Dave Rensin’s account of the method, a crashed AI session helped expose the problem: restarting the work meant reconstructing the project’s intent. Keeping that intent in a document made it possible to resume without relying on the lost chat.
That does not make a fresh AI reviewer an objective expert. It may still share the same underlying model’s weaknesses, misunderstand a file or miss a failure case. Separating conversations removes one source of shared assumptions; it does not remove every source of error.
What the plugin actually does
The core workflow begins with brainstorming, moves into a written draft and then sends that draft through review. The developer participates throughout. The plugin does not make the architectural decisions on the developer’s behalf.
Brainstorming tests the idea and records an agreed direction in concepts_brief.md. Drafting produces a design document section by section, with pauses for feedback. For the webhook example, a useful design would need to settle the failed-delivery and duplicate-event questions rather than leave the implementation agent to guess.
The current review pipeline has five stages. It first asks whether the design is unnecessarily complicated or duplicates something the project already has. A comprehension check then tests whether a fresh reader can understand it. Technical critics examine the design, another simplicity pass checks the additions, and an implementation-readiness review asks whether enough detail exists to build it.
A reviewer needs enough detail to build the feature, not just understand its purpose. The plan should identify the files to change and the interfaces and tests needed to carry it out.
The developer chooses which findings to accept, resolves the remaining trade-offs and approves the plan. Only then does the implementation work begin. EGM’s design-stage instructions say not to change production code, but its architecture document explicitly calls this a convention, not an enforced security boundary.
The plugin also includes tools to build documentation for an existing codebase and report “drift”, where the code and its design documents no longer agree. A drift report does not decide which is correct. A person has to choose whether to update the software or the document.
“Zero context” is not zero information
A goldfish needs the specification and the source files required to assess it. “Zero context” refers to the earlier conversation, not to reviewing a system without evidence.
There is also an important implementation limit. The project’s host integration documentation distinguishes isolation from earlier chat turns from exclusion of instruction files. A host is the application running the plugin, such as Claude Code or Codex. It can add project-wide instructions to a new session even when that session does not inherit the parent conversation.
The docs record reports of project instructions reaching Codex and Antigravity reviewers. Claude’s reviewer adapters explicitly request exclusion of CLAUDE.md; the documented Codex and Antigravity checks do not establish equivalent instruction-file exclusion. That difference matters if an instruction file contains the very assumption the reviewer is supposed to challenge.
If the host cannot create fresh subagents, EGM stops the review rather than presenting a review in the same conversation as equivalent. Installing a plugin successfully is therefore not enough to prove the review works as intended.
Where it runs, and what “supported” means
EGM is available under the Apache 2.0 licence, with no separate hosted EGM account required. It is published in Google’s GitHub organisation, but the repository says it is not an officially supported Google product. Its test record separates installation, workflow execution and reviewer-isolation evidence. CLI means command-line interface: the version used through a terminal.
| Environment | Documented position | Practical limit |
|---|---|---|
| Claude Code | Plugin installation, full-workflow tests and tests of reviewer isolation are documented | Use the complete plugin, including reviewer instructions and document standards |
| Codex CLI | Installation and workflow tests are recorded, including isolation from parent chat turns | Do not assume project and user instruction files are excluded from reviewers |
| Antigravity CLI | Full-workflow tests and a test of isolation from earlier chat turns are recorded | Instruction-file exclusion is not established by those tests |
| ChatGPT and Codex desktop; cloud workspaces | A desktop plugin listing and workspace installation routes exist | Cloud review evidence remains partial; the current session must expose fresh-subagent tools |
| Standalone Antigravity IDE | Not supported for the Goldfish workflow | The documented runtime lacks the fresh-context subagent tool required for review |
These distinctions come from the project’s environment guide, not a claim that every installation has been independently tested. The standalone Antigravity IDE is also distinct from the Antigravity App and its Visual Studio Code extension, which have separate installation instructions.
For Claude Code, the repository gives the following installation commands. Other hosts use their own setup routes; copying only the skills folder leaves out files the workflow needs.
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egmThe cost is in the agent work, not a plugin fee
Downloading an open-source plugin does not make its model calls free. EGM runs through the host account, so repeated document reads and multiple reviewers consume that account’s allowance or paid usage. The amount depends on the model, document length, caching and how many review rounds run. There is no useful fixed price for “an EGM review”.
Independent developer Mashhood Rastgar compared EGM, the Superpowers agent workflow and a plain Claude Code agent on a tool that merges bank transaction files. All three passed the 18 hidden acceptance tests. The more revealing results appeared when he tried cases outside those tests.
EGM’s design process specified a refusal to overwrite an input file, preventing a mistake the other two implementations allowed. But EGM and the plain agent both lost valid transaction rows after a malformed quoted field. Superpowers caught that problem in a final review of the complete code changes.
For that run, Rastgar reported roughly 100 minutes and US$31 for EGM, compared with 28 minutes and US$6.46 for Superpowers and five minutes and US$0.71 for the plain agent. These are reconstructed model costs for particular runs, not prices for the plugins. Different subagents used different model tiers.
He repeated the comparison on stock-taking and invoicing tools. His published assessment favoured EGM’s design documents and Superpowers’ resulting code, but neither workflow consistently caught the data-loss cases. The experiment repository also records limitations: one run per workflow per problem, an AI stakeholder that sometimes volunteered hidden requirements, and automated judges whose mistakes needed correction. Output tokens were undercounted, making the cost figures lower bounds.
That is useful evidence of trade-offs, not a general ranking of the products. A document review and a running-code check catch different things. More review rounds can produce more wording changes without discovering the bug that loses someone’s data.
Fresh reviewers are not a privacy guarantee
The project’s data-handling notice says the runtime operates no hosted service, database, analytics client or telemetry endpoint. It does not require separate EGM credentials. But the host assistant still processes the prompts, files and agent activity under its own terms and account settings.
A new reviewer conversation does not mean the model runs locally, that a provider retains nothing or that confidential code stays on the computer. Review documents can also repeat confidential material into local reports or files stored by the host. EGM does not set the host’s retention policy, and local output files persist until deleted.
The bootstrap workflow can run local discovery commands and an existing project formatter; those commands may contact outside services. The plugin installs no permission-approval hook that replaces the host’s controls.
Adoption and review overhead
Google’s announcement says thousands of its engineers use the plugin. That describes internal adoption, not a controlled measure of how much faster or safer an outside team will be. The project’s own FAQ warns that extensive documentation can become a review bottleneck and that small, well-understood changes may not justify the process.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
