EGM: Google's plugin that lets memory-less agents tear into your design first
Google open-sourced the design-review plugin it says thousands of its engineers use daily. Goldfish reviewers with no chat history read your spec cold
- Who
- Google's open-source plugin, Apache 2.0. Per the README thousands of Google engineers use it; not an officially supported product.
- When
- Repository created on 2026-09-13, announced on the Google Open Source Blog on 2026-10-22, first tagged release v0.1.0 on 2026-10-27.
- Where
- A plugin for Antigravity CLI, Claude Code and Codex; docs also cover Cowork, ChatGPT Work, omp and Muse. Runs in your agent with subagents.
What was hard
Hand a coding agent one feature and the first hours fly, then a few days later you are stuck in a long thread where fixing one bug breaks two others. Writing the design first did not help much either: the same session that wrote it reviewed it, and believed the same missing assumptions.
What it does about it
EGM splits the work into two roles. The Elephant, a session holding the whole project context, drafts the design document and writes no code at all. Goldfish subagents with no conversation history then read that document as a newcomer would and flag hidden assumptions and empty explanations. You resolve the trade-offs and approve the design, and only then ask the agent to implement it.
What people made with it
A short report from someone who turned EGM on in Antigravity, pointed it at their own projects, and got a long list of fixes from its audit. They note no affiliation.
so i told it to look at my projects and it found a lot to fix with its audit
How it runs inside
- 1
You → Elephant (Claude Code session)
/egm-brainstorm one lineThe elephant drills the problem with open questions, runs six goldfish lenses on it, and writes concepts_brief.mdskills/egm-brainstorm/SKILL.md - 2
Elephant (Claude Code session) → egm-draft
concepts_brief.mdTurns the brief into design_draft.md section by section, with exactly one approval question after eachskills/egm-draft/SKILL.md - 3
You → egm-review
/egm-review design_draft.mdPicks a review preset and splits the doc at the '## Detailed Implementation Plan' heading into two rangesskills/egm-review/SKILL.md - 4
egm-review → Goldfish subagents
persona file + doc path, no chat historyA fresh subagent per stage; the design critic hunts unbacked assumptions and unmitigated risks and may not editagents/design_critic.md - 5
Goldfish subagents → egm-review
Blocker / Major / Nit findingsThe comprehension reviewer paraphrases first; a section it cannot paraphrase is a Blocker. Reports go to .md filesagents/comprehension.md - 6
egm-review → Elephant (Claude Code session)
egm-resolve: contest, fix, escalateThe elephant labels each finding CONTEST, PROPOSE FIX, ESCALATE or DECLINE and applies approved edits to the draftskills/egm-resolve/SKILL.md - 7
Elephant (Claude Code session) → Approved design doc
approved design under docs/designs/You approve, the agent builds to the document, and /egm-drift-report later compares the document with the codedocs/getting_started.md
What goes in, what comes out
Is it for you
Good for
- Building a feature that needs design decisions with a coding agent
- Working alone with nobody to review your design
- Claude Code, Codex or Antigravity CLI sessions that can spawn subagents
- Long agent sessions where unstated assumptions pile up
Not for
- The README says to use your normal coding workflow for a small, clear bug fix
- Goldfish reviews need fresh-context subagents; where the host cannot create them, the review stops
- A review does not replace tests or human judgment
- v0.1.0 is the first tagged release, and its notes say not to read it as a stable release
Reviewed by Devalley’s editor before publishing. How the column is made