Elephant-Goldfish Model (Agent Skill) news update

Google published Elephant-Goldfish Model, an open-source plugin for AI coding assistants under the Apache 2.0 license, according to github.com's report. Released as an open-source project rather than an official Google product, EGM implements specification workflows inspired by Dave Rensin's essay, dividing development between an elephant holding conversational context and goldfish reviewers examining drafts with fresh eyes.

This report details installation methods, command prefixes, agent requirements, and operational skills across supported AI assistant environments. The release serves software engineers and systems architects using terminal or desktop coding tools who require structured design reviews before writing code. Through modular skills for brainstorming, drafting, reviewing, bootstrapping codebases, and detecting drift, EGM formalizes how developers guide autonomous AI architecture.

Elephant-Goldfish Model (Agent Skill) news: what actually changed

The release introduces an open-source plugin operating across multiple AI coding assistants, providing standardized skills to draft and evaluate software specifications. In desktop environments, users install the published listing through ChatGPT and Codex desktop applications. For terminal workflows, Google published installation commands for Antigravity CLI, Claude Code, and Codex CLI. Engineers running Antigravity CLI execute agy plugin install https://github.com/google/elephant-goldfish-model, while Claude Code users add google/elephant-goldfish-model and install egm@egm. Codex CLI uses matching commands, and Antigravity IDE users access the /plugin command.

Beyond installation packaging, the release alters command handling between CLI versions. On Antigravity CLI version 1.4.10, issue #220 established completion from unprefixed names to the required /egm: prefix, whereas Antigravity CLI version 1.4.8 lacked this under issue #196, making prefixes like /egm:egm-brainstorm necessary. In Claude Code, the /egm: prefix remains mandatory, while Codex uses the skill picker. After installation, developers must start a new session to load reviewer prompts and document standards, with setup guidance provided for Claude Cowork and ChatGPT Work environments.

Elephant-Goldfish Model (Agent Skill) news: how it works

The system architecture partitions responsibilities across three distinct roles: the elephant, goldfish reviewers, and the human developer. The elephant retains project context, requirements, and earlier conversational decisions, exploring feature requirements and drafting the specification document. In contrast, goldfish reviewers examine the draft and referenced source material without prior conversational context, finding unstated assumptions, unclear requirements, and implementation risks that context bias obscured. The human user manages project trade-offs, evaluates findings, and approves the architecture before deciding what to build and implement in production code.

A design cycle begins by exploring a feature such as webhook delivery with retry failures or duplicate payloads using /egm-brainstorm. The discussion captures agreed directions in concepts_brief.md before the /egm-draft command translates that brief into design_draft.md, pausing after each section for interactive chat feedback. The /egm-review command then invokes fresh subagents to inspect clarity, codebase fit, and implementation planning. If the host environment cannot spawn fresh subagents, EGM halts review execution immediately. Approved specifications are saved under docs/designs/ prior to implementation.

Elephant-Goldfish Model (Agent Skill) news: why it matters

Standard conversational coding workflows frequently suffer from context drift and confirmation bias, where an assistant uncritically agrees with developer proposals after long chat sessions. EGM addresses this failure mode by grounding its workflow in Dave Rensin's original essay on two-tier context evaluation. Separating specification drafting from review forces technical designs through objective validation gates before code generation starts. By generating an explicit concepts_brief.md and an iterative design_draft.md, the framework ensures engineering teams evaluate edge cases in writing rather than discovering architectural shortcomings midway through coding.

The model also introduces maintenance capabilities that protect software health across evolving repositories. Developers can invoke /egm-review-loop to run configured reviewer groups and custom standards against non-software documents. For legacy codebases lacking architecture records, the bootstrap skill creates documentation directly from existing source files. Additionally, the drift report skill compares checked-in design documents against current code to identify divergence. Distributed under the Apache 2.0 license, the plugin gives developers open access to reviewer prompts and document standards while maintaining an accessibility statement documenting known barriers.

Elephant-Goldfish Model (Agent Skill) news: who is affected

The release directly affects developers using supported AI coding tools including Claude Code, Codex CLI, Antigravity CLI, and desktop applications for ChatGPT and Codex. Engineers who rely on automated agents for architectural drafting gain modular skills such as brainstorm, draft, resolve, bootstrap, five-stage review, review loop, and drift report. However, users in environments lacking subagent support cannot run reviews, as EGM automatically stops when fresh subagents cannot be spawned. Developers must verify that their coding host supports parallel agent sessions before integrating the review pipeline.

Development teams and engineering managers are also impacted by verification boundaries established in the documentation. Google explicitly states that EGM is not an officially supported Google product, shifting operational responsibility for trade-offs and approvals to the human developer. Automated reviews do not replace unit testing, integration tests, or human judgment. Instead, human engineers must review findings, decide which architectural trade-offs to accept, and approve specifications before writing production code. Teams adopting the tool must establish review protocols around docs/designs/ so automated feedback supplements human oversight.

Elephant-Goldfish Model (Agent Skill) news: what to watch

Technical teams adopting EGM should observe how host compatibility and command handling evolve across agent releases. Users operating Antigravity CLI should monitor version-specific behaviors, noting that version 1.4.10 completes unprefixed commands to /egm: under issue #220, while version 1.4.8 requires explicit prefixes under issue #196. Practitioners should also track installation guides for collaborative platforms like Claude Cowork and ChatGPT Work. As desktop extensions expand agent features, broader support for fresh subagent context creation will determine which developer platforms can run five-stage goldfish reviews natively.

Organizations should also watch how community contributions shape the open-source repository under Apache 2.0. Developers can submit pull requests, log bugs through issue tracking, and participate in discussions to refine prompt standards and accessibility features. Teams implementing software architecture will monitor the practical performance of the drift report skill in active codebases to evaluate how automated diffs catch architectural decay. As developers test modular skills for brainstorming, drafting, and bootstrap documentation, community feedback will show whether isolating review context provides a dependable foundation for software engineering.