Today for AI
HOT RADAR
agentHEAT 16.2°

AI CURATED HOT STORY · 11/8/2026

Google Open-Sources EGM Plugin to Make Coding Agents Write Specs First

1 reports archived1 independent sourcesupdated 11/8/2026, 18:48:54
Synthesis & Latest Updates
1 Sources Cross-Validated

Google has open-sourced the Elephant-Goldfish Model (EGM) plugin under Apache 2.0, requiring coding agents to draft and defend a design spec before writing code. The core idea is context splitting: a long-lived "Elephant" session holds project history and drafts the design doc, while fresh "Goldfish" subagents read only that doc and referenced files, avoiding inherited blind spots. The review runs five stages; in Google's tests Codex 0.163.0 and Antigravity CLI completed 15-16 reviewers end to end, though reviews consume the user's own model quota and issue #237 reports AGENTS.md leaking into some CLI reviewers.

LATEST/Google has open-sourced the Elephant-Goldfish Model (EGM) plugin under Apache 2.0, requiring coding agents to draft and defend a design spec before writing code. The core idea is context splitting: a long-lived "Elephant" session holds project history and drafts the design doc, while fresh "Goldfish" subagents read only that doc and referenced files, avoiding inherited blind spots. The review runs five stages; in Google's tests Codex 0.163.0 and Antigravity CLI completed 15-16 reviewers end to end, though reviews consume the user's own model quota and issue #237 reports AGENTS.md leaking into some CLI reviewers.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. Marktechpost (@Marktechpost)T2·72 pts

    Google Open-Sources EGM Plugin to Make Coding Agents Write Specs First

    • Core mechanism is context splitting: a long-lived "Elephant" drafts the design doc while fresh "Goldfish" subagents read only the doc and referenced files, avoiding main-session blind spots
    • Review has five stages: zoom out, comprehension test, critic review, second zoom out, and implementation-readiness check
    • Ships under Apache 2.0 with 7 skills and 9 reviewer personas, works in Antigravity, Claude Code 2.1.271+, Codex and ChatGPT, but each review burns the user's own quota