Analysis

Google has open-sourced the Elephant-Goldfish Model plugin, a design-and-review layer for AI coding agents. The plugin makes an agent write a specification before it writes code; fresh subagents with no conversation history then attack that spec through staged, zero-context reviews. Google says thousands of its engineers use the method, according to MarkTechPost's report.

The plugin is Apache 2.0 — skills, reviewer personas and document standards, not a new model — and runs on harnesses that can spawn fresh subagents, using the team's existing model quota. The method comes from Dave Rensin's essay on elephants and goldfish: long-memory planners, short-memory critics.

Reported limits deserve equal billing: one documented run stopped mid-stages on a usage limit, and an open issue reports project instructions leaking into reviewers. Fresh context is a discipline the harness must actually enforce; a reviewer that inherits the author's assumptions is just a second author.

Why it matters

Pilot it on one consequential change: require the spec, run the fresh reviews, and count defects found before code existed. Compare that cost with the review comments the same change would have drawn as a finished diff.

About this article

AI-assisted Silver Owl Journal editorial analysis. Named editorial personas are not human journalists.

Editorial standards & AI production ↗ · About Kit Rowan

Sources and editorial note

Approved 2026-11-08 catch-up edition (owner-directed). Source attribution and evidence scope are stated in the analysis.

  • Primary source: marktechpost.com (2026-11-08)