AI NewsDev toolsAnnouncement

Google released EGM, a spec-driven plugin for Claude Code, Codex and Antigravity

Google published a spec-driven plugin that runs across Antigravity, Claude Code, Codex and Claude Cowork. The repo had 984 stars and ranked 24th on Trendshift.

AI News

Editorial2 min read

LinkedInX
GitHub social card for google/elephant-goldfish-model

Image: GitHub

Why it mattersA team using a coding agent gets a shared spec step before it writes code. The spec is reviewed by many small agents, each with no history.

Google published Elephant-Goldfish Model, abbreviated EGM, on GitHub. The repo had 984 stars and 91 forks on 25 October 2026, and ranked 24th on Trendshift's daily list. It uses Apache 2.0 and runs on Antigravity, Claude Code, Codex and Claude Cowork. The README says it is not an officially supported Google product. The methodology comes from Dave Rensin, whose Medium essay the project links to.

The project solves one problem. An AI coding agent writes code before anyone agrees what to build. The pull request then has to be rolled back. EGM adds a short spec step before the code. The agent itself runs it.

Two animals, two roles

The README names two sides. The Elephant is a deep-context agent. It loads the system history, the architecture notes, the component boundaries and the user journeys. It runs structured brainstorming and drafts a specification under docs/designs. It carries the whole picture.

The Goldfish is a swarm of short-lived subagents. Each one reads only the draft spec and the live source code. Each one has no shared history. The README names seven personas in agents/: Design Critic, Naive Reader, Tech Lead, Codebase Auditor, Structural Validator, Style Editor and Zoom Out Architect. Each persona reads the draft from a different angle. The author Dave Rensin writes that the human engineer still owns the architectural decision, the trade-offs and the final outcome.

Seven skills a developer runs

The seven skills are slash-commands inside the agent, each with an egm prefix. Brainstorm runs an ideation session with a Goldfish swarm that shoots down early concepts. Draft is a section-by-section compiler of a design doc. Review runs a multi-stage pipeline: Zoom Out, Comprehension, Critic, Readiness. Review-loop dispatches review subagents in parallel or in a hierarchy. Resolve is interactive triage of the findings. Drift-report scans the current code for divergence from a checked-in spec. Bootstrap is a recursive codebase documentation generator.

Where EGM sits

Spec-driven development is not new in the AI coding tool space. The field has an Akka tool, a Pulse study, and a long InfoQ piece on when spec-driven development pays off. What is new in EGM is a shared vocabulary. The plugin works across the big three agents on one install. The deep-context writer and the stateless reviewer are the single idea worth copying.

The repo is six weeks old. 984 stars and 91 forks are small numbers against the top of GitHub. The Trendshift ranking is a daily movement number. For a team that uses an AI agent on real repositories, the test is one afternoon. Draft a design doc for a feature already in flight. Run the Goldfish swarm. If the swarm finds a hole the engineer did not see, EGM earns a slot in the workflow. If not, the agent still writes code.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project