Resume
← Blog

2026.11.01 · 4 min · ai · agents · claude-code · tooling

Review your plan with an AI that remembers nothing.

Google's Elephant-Goldfish Model splits AI coding into a long design chat (the Elephant) and fresh, memoryless reviewers (the Goldfish) that get only the design doc. Whatever a Goldfish has to guess is a gap in the doc, fixed before any code. How it works, the prompts to try it without the plugin, and what it doesn't replace. As of 2026.11.

After a long chat with your coding agent, the plan feels clear. It feels clear to the agent too. That's the problem: you both remember the whole conversation, so neither of you notices what the design doc never actually says. The next teammate, a fresh session, or the same agent after a crash gets only the doc.

Watch the 60-second version on Instagram.

The Elephant-Goldfish Model (EGM) is a workflow built around that gap. It isn't an AI model, despite the name. Dave Rensin coined it at Google, and Google open-sourced it as an Apache 2.0 plugin on 2026.10.22. Google says "thousands of Googlers" use it daily to ship production code, though it doesn't publish measured results.

Two kinds of session

The Elephant is one long-running, context-rich session. It writes no code. It only handles design, intent and decisions, and its output is a Markdown design doc.

The Goldfish are stateless, zero-context sessions. Each one gets the doc and nothing else, and reviews it.

ELEPHANT · LONG CHATGOLDFISH · NO MEMORYdesign.mdONLY THE DOC CROSSES THE WALLFIG. 01 / ONLY THE DOC CROSSES

The chat history never reaches the Goldfish. That's the whole design. Forgetting usually reads as a weakness in an AI session; here it's the point. A Goldfish can't fill gaps from memory, so it shows you exactly what the doc leaves out.

Google frames the problem it targets bluntly: AI speeds up bugs and "slop" too, so teams risk "auto-scaling our mistakes". The project's motto is "Design is the new code."

The review: a guess is a gap

The plugin has three commands: egm-brainstorm, then egm-draft (a Markdown doc with the problem, the plan, alternatives and a file-level implementation plan), then egm-review. The launch post describes three checks:

  1. Comprehension. A Goldfish has to explain the design back.
  2. Antagonistic. Skeptical Goldfish hunt for edge cases.
  3. Implementation readiness. A final go/no-go confirms file paths, data structures and signatures.
design.md## PROBLEM## PLAN## ALTERNATIVES## FILESuse the usual cache?01 COMPREHENSION02 ANTAGONISTIC03 READINESSCODE · OPENA GUESS IS A GAP · FIXED BEFORE CODEFIG. 02 / A GUESS IS A GAP

If a Goldfish reads "use the usual cache", it has no idea which cache. That line goes back to the Elephant, gets rewritten, and the review runs again. Only when the Goldfish can explain the plan without guessing does code start. The repo has grown since launch: it now runs 5 review stages with 7 reviewer personas.

Try it without the plugin

You don't need anything installed to get most of the value. Two chats and one file:

two-sessions.txt
# Session 1 (Elephant): design only
"Don't write code. Interview me, then write design.md."
# Session 2 (Goldfish): brand-new chat
"Here is design.md and nothing else. Explain back what will be
built and which files change. List anything you had to guess."

The list of guesses is the output that matters. Each item is a sentence missing from the doc.

If you want the full workflow, the plugin installs into Claude Code, Codex and Antigravity (early launch coverage said Antigravity only, with the others "upcoming"; the repo now lists all three):

terminal
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

It needs an agent that can spawn fresh subagents. Without that, in the README's words, "EGM stops the review".

What people get wrong

  • "Google released a new AI model called EGM." No. It's a workflow plus prompts and skills.
  • "I'll paste the chat into the Goldfish so it has context." That defeats the test. The Goldfish gets the doc and nothing else.
  • "The doc passed review, so the code is right." The repo says it plainly: "A review does not replace tests or human judgment."
  • "It reviews security too." Security isn't a built-in review category. Add a reviewer persona for it yourself.
  • "Use it for everything." The repo says to skip it for small, clear fixes. It's overhead that pays off on features big enough to need a design doc.

On a client project I'd run the Goldfish pass on any feature that touches more than one service, and I'd add the security persona before the first review, not after the first incident.

The short version

  1. The Elephant is one long design chat that writes design.md and no code.
  2. A Goldfish is a fresh session that gets only the doc.
  3. Anything the Goldfish has to guess is a gap; fix it in the doc before coding.
  4. You can try it with two chats; the plugin adds stages and personas, and needs subagents.
  5. It doesn't replace tests, and security needs a persona you add yourself.

Sources: Google Open Source Blog, Introducing the Elephant-Goldfish Model plugin · google/elephant-goldfish-model · EGM architecture doc · Google Research, Elephants, Goldfish and the New Golden Age of Software Engineering · Help Net Security · UMD CMSC389A, The Elephant-Goldfish Model · Open Source For You