Elephant-Goldfish Model (EGM)
Draft specs with a deep-context 'Elephant' agent, then stress-test the design with zero-context 'Goldfish' reviewers to expose unstated assumptions before you write code.
- Source repo
- google/elephant-goldfish-model
- Stars
- ★ 5.6k
- Last updated
- today
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 75/100 · Good
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude CodeChatGPT · Claude.ai (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An engineer adding a feature that needs design decisions—e.g., webhook delivery and how to handle failed or duplicate deliveries—starting with brainstorming to clarify trade-offs
- Not a fit if
- Coding-agent environments that cannot spawn fresh subagents
- Developers fixing small, clear bugs with no design decisions
- Teams not using Antigravity, Claude Code, or Codex
- Source review
- 75/100 · Good
What does this agent do, and when should you use it?
EGM is an open-source plugin from Google that adds a set of design-and-review skills to your coding agent; per the README, thousands of Google engineers use it. The core idea is role separation: the 'Elephant' holds your project context and drafts the specification, while 'Goldfish' reviewers examine the draft and referenced source material in fresh subagent contexts without the earlier conversation, exposing unstated assumptions and blind spots. You make the architectural trade-offs and approve the spec before implementation. The plugin installs via the plugin mechanisms of Antigravity, Claude Code, or Codex and is invoked through slash commands such as /egm-brainstorm, /egm-draft, and /egm-review, or via Codex's skill picker. The workflow produces documents like concepts_brief.md and design_draft.md, with review covering the design's clarity, fit with the code, and implementation plan. Goldfish reviews require fresh subagent contexts—EGM stops the review if your agent cannot create them—and reviews do not replace tests or human judgment. It is released under Apache 2.0 and is not an officially supported Google product.
EGM installs as a set of agent skills. /egm-brainstorm has the context agent discuss requirements and trade-offs with you while fresh reviewers challenge the proposal, saving the agreed direction in concepts_brief.md; /egm-draft turns the brief into design_draft.md, pausing after each section for your feedback; /egm-review runs fresh reviewers over the draft, checking the design itself, its clarity, its fit with the code, and its implementation plan. Additional skills: /egm-resolve works through review findings, egm-bootstrap creates or updates documentation for an existing codebase, egm-review-loop runs a configured group of reviewers for a specific task, and egm-drift-report compares checked-in documents with current source code. Once you approve the design, you ask your coding agent to implement it.
- An engineer adding a feature that needs design decisions—e.g., webhook delivery and how to handle failed or duplicate deliveries—starting with brainstorming to clarify trade-offs
- Teams wanting a zero-context stress test of a design document before implementation to surface unstated assumptions
- Developers onboarding onto an unfamiliar codebase who use Bootstrap to generate or update project documentation
- Maintainers running Drift report to check whether checked-in design docs have drifted from current code
- Developers who want the same design workflow across multiple supported agents (Antigravity, Claude Code, Codex)
How do you install or deploy this agent?
Pick the install path for your coding agent. Antigravity:
agy plugin install https://github.com/google/elephant-goldfish-modelClaude Code:
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egmCodex:
codex plugin marketplace add google/elephant-goldfish-model
codex plugin add egm@egmStart a new session after installing. For Claude Cowork and ChatGPT Work, follow the setup guide (docs/harnesses.md) in the repository.
How do you use this agent?
Open your project in your coding agent and choose a feature that needs design decisions. In Antigravity and Claude Code, add the /egm: prefix, e.g. /egm:egm-brainstorm; in Codex, select the corresponding EGM skill from the skill picker. Typical flow:
/egm-brainstorm Add webhook delivery to this project./egm-draft/egm-review design_draft.mdBrainstorming and review require a session that can create fresh subagents. The review returns the draft path; you choose which findings to address, and once you approve the design, ask your coding agent to implement it. See docs/getting_started.md for the full process, including saving approved specs under docs/designs/.
What are this agent's strengths and limitations?
- Distinctive zero-context 'Goldfish' reviewer design: examining the draft without the earlier conversation reliably surfaces unstated assumptions and blind spots
- Works as one consistent set of plugin skills across three major coding agents (Antigravity, Claude Code, Codex), each a one- or two-command install
- Skills are usable independently—no need to run the full brainstorm→draft→review sequence—and include extras like Bootstrap and Drift report
- Backed by real internal use by thousands of Google engineers, with an architecture doc and testing records for auditability
- Depends on the host agent being able to create fresh subagent contexts; EGM stops the review otherwise
- Fully documented only for Antigravity, Claude Code, and Codex (with partial support for Claude Cowork and ChatGPT Work); other agents require adaptation
- Scoped to pre-implementation spec work; small, clear bug fixes should use your normal coding workflow instead
- Explicitly does not replace tests or human judgment, and is not an officially supported Google product
How does this agent compare with similar options?
The README cites Dave Rensin's essay 'Elephants, Goldfish, and the New Golden Age of Software Engineering' as the origin of the ideas behind EGM; the source names no directly comparable alternative products.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Elephant-Goldfish Model (EGM) This agent | 75 · Good | Agent plugin / skillFree + model costs | ★ 5.6k | today | Python | Codex · Claude Code |
| Finding-Unknowns Skills | 76 · Good | Agent plugin / skillFree | ★ 358 | 1mo ago | Python | Claude Code · Claude.ai |
| Smart Ralph | 73 · Some gaps | Agent plugin / skillFree + model costs | ★ 563 | 1mo ago | Shell | Codex · Claude Code |
| JS Mastery Engineering Workflow Skills | 58 · Major gaps | Agent plugin / skillFree + model costs | ★ 1.6k | 2mo ago | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
privacy.md explicitly states no hosted service, no telemetry, no credentials required; bootstrap can run local discovery commands and a user-project formatter, with users directed to authorize via host permission controls — least privilege and user confirmation are documented but ultimately depend on the host, so not full marks. Data flow transparency is excellent, covering host retention limits and output echoing inputs. Rollback is weak: artifacts are markdown written directly into the workspace with no retention or recovery mechanism beyond manual deletion. Source attribution is strong: full Apache-2.0 text, Google organization, attribution to Dave Rensin's original essay, and an explicit 'not an officially supported Google product' notice.
Docs and tests are largely consistent, but there is residual tension: README says the review runs in 'five stages', getting_started describes a '6-lens Goldfish swarm', and the integration test enumerates six stage reports and 'six brainstorm and nine stage reviewers'; the cross-references are not fully reconciled within the provided files. No runtime dependencies; five host harnesses supported with documented status. Failure behavior is specified in prose (review stops if subagents unavailable; tests demand explicit failure when tools are missing) but no executed evidence.
Use-case boundaries are crisp: EGM for features needing design decisions, normal workflow for small clear bug fixes; the FAQ honestly answers 'it depends' on velocity. Capability boundaries are excellent — repeated statements that reviews do not replace tests or human judgment, and that zero-context review is 'sometimes, but not always' realistic. Trigger precision is moderate: seven skills are well differentiated but multi-harness prefix differences rely on documentation. Environment fit spans five harnesses, though Claude Cowork/ChatGPT Work details live in an external guide not included here.
Excellent information architecture: README, quickstart, harnesses, FAQ, privacy, and TESTING.md each have clear roles, with per-skill SKILL.md files. Install instructions give concrete commands per host. egm-* naming is consistent, with minor prefix-description discrepancies between README and getting_started. FAQ and known-limitations content is thorough and honest. Full Apache-2.0 license present. Deductions: no CHANGELOG or versioning convention is visible; maintenance responsibility rests on the Google org and contribution channels, but the 'not an officially supported product' notice leaves long-term commitment unclear.
Deliverables are well defined (design doc, implementation checklist, context README hierarchy) and the draft flow pauses per section for feedback — good output usability. Marginal value: fresh-context Goldfish review is a clear methodological contribution, but effectiveness claims such as 'thousands of Google engineers use EGM' cannot be verified within the repository. Cost-benefit is addressed honestly in the FAQ (design review can slow small changes; prioritize) — candid, but no quantitative benefit evidence is provided.
Claims map to documents: architecture, testing, and privacy docs cross-reference; opt-in integration tests concretely verify reviewer isolation (marker non-leakage, non-forked spawns, denied persona reads) — strong traceability evidence, though none were executed in this static review. Fact/inference separation is good: the FAQ uses hedged, opinion-labeled language. Deductions: 'thousands of Google engineers use EGM' is an unverifiable external claim, and TESTING.md and some referenced docs exist only as links outside the provided evidence.
- This is a static source review; no skills or integration tests were executed. Key claims such as reviewer isolation rely on opt-in tests and should be verified independently.
- EGM writes design docs and READMEs directly into your workspace with no automatic rollback; put the workspace under version control before use.
- Data flow is governed by the host agent (Claude Code/Codex, etc.); EGM does not control host transcript or retention policies — provide only the files needed.
- Bootstrap can execute locally configured commands such as formatters, which may contact package registries; authorize them individually through host permission controls.
- Claims like 'thousands of Google engineers use EGM' and productivity benefits cannot be verified from the repository; the plugin is explicitly not an officially supported Google product.