Elephant-Goldfish plugin: spec review for coding agents

What changed

Google has open-sourced the Elephant-Goldfish Model plugin under the Apache 2.0 license. The tool functions as a design-and-review layer for AI coding agents rather than a new language model. It establishes a structured workflow where an agent with full project context, called the Elephant, drafts a specification in Markdown before writing any code. Once the specification is ready, stateless subagents called Goldfish reviewers examine the draft and referenced source files without access to earlier conversation history to identify unstated assumptions and overlooked risks. The human then resolves trade-offs and approves the design before implementation begins.

Google reports that thousands of its engineers already use this approach internally. The plugin provides a collection of skills, reviewer personas, and document standards designed to catch flawed assumptions in specifications rather than during pull request review, when fixes become costlier and affect larger portions of the codebase.

The workflow includes multiple commands: brainstorm clarifies ideas using six design lenses covering YAGNI, token cost, UX, migration, observability, and technical debt; draft builds a design document section by section; review runs the Goldfish review process; and drift-report compares checked-in specifications with current code, with the documentation recommending weekly runs.

Why it matters

Teams using agent-based development face a fundamental problem: as AI generates code faster than humans can thoroughly review, the specification becomes more critical than the code itself. Traditional code review cycles catch issues late, during pull requests, when fixes require more effort and affect larger portions of the system. The Elephant-Goldfish workflow validates design decisions earlier by enforcing independent specification review before implementation begins.

This approach matters most for engineering teams relying on coding agents. For product teams, the design-first discipline reduces architectural rework cycles and clarifies handoffs between planning and implementation phases. The fresh-context Goldfish reviewers represent a concrete change to the review process: they see only the specification and referenced files, not the full conversation history that shaped the design. This isolation surfaces clarity problems in the documentation itself rather than relying on shared conversational context.

The plugin works across multiple agent platforms, including Antigravity CLI and App, Claude Code version 2.1.271 or later, Claude Cowork, Codex, and ChatGPT. This multi-platform support matters because teams already using different tools can adopt the workflow without switching vendors. However, the plugin requires agents capable of spawning fresh subagents. If a platform cannot create isolated subagent contexts, the review process stops entirely.

Compared with similar tools, Spec Kit and Kiro center on human review between development phases, while Superpowers isolates subagents mainly during implementation. The Elephant-Goldfish approach stands out in reviewing specifications with zero-context agents before any code is generated.

What to test

Before adopting EGM in production workflows, teams should verify several claims and understand known limitations:

• Fresh-context isolation: Test whether your chosen harness genuinely creates isolated subagent contexts for Goldfish reviewers. The source notes that reviewer isolation still leaks host instruction files on some CLIs, and AGENTS.md files have leaked into Codex reviewers in at least one case. Run a test review with sensitive design information and check what context the reviewers actually receive.

• Token consumption: The plugin runs five sequential review stages with multiple reviewer personas per stage. Measure token usage against your quota and baseline agent costs. The source flags heavy token consumption as a known issue, making this a critical cost consideration.

• Harness compatibility: The source provides concrete test data showing Codex 0.163.0 passed automated tests with 15 native reviewers in 573.13 seconds, and Codex CLI completed the full suite in 482.30 seconds. Test your specific agent version and harness to confirm comparable behavior. Claude Code requires version 2.1.271 or later; Claude Cowork has reported usage limit issues that stopped reviews before completion.

• Reviewer tool constraints: Claude Code reviewers receive only Read, Grep, and Glob tools, and CLAUDE.md files are excluded from their context. Verify these constraints do not prevent reviewers from checking the code paths and data structures relevant to your design decisions.

• End-to-end run stability: Antigravity CLI 1.4.5 passed testing with 16 native reviewers running stages in order. Your setup may differ. Run a complete brainstorm-draft-review cycle on a non-critical feature and observe whether all five review stages complete without timeouts or agent errors.

• Drift reporting accuracy: The plugin offers drift-report functionality to compare specifications with current code over time. Test whether drift detection accurately identifies divergence between your approved design document and actual implementation.

The conclusion

The Elephant-Goldfish coding agent plugin represents a concrete shift in how teams structure agent-driven development: specifications first, reviews with fresh context, code second. For teams already using coding agents, this workflow reduces late-stage architectural failures and clarifies handoffs between planning and implementation phases. The multi-platform support and open-source licensing lower adoption barriers.

The real costs are token consumption and verification complexity. Heavy-use teams need to measure baseline impact against their model quota before deployment. The plugin’s reliance on fresh-context agents means compatibility requirements are strict; if your harness cannot spawn isolated subagents, EGM stops functioning. The documented issue of reviewer isolation leakage on some CLIs indicates this is still a maturing tool.

What to watch: whether Google publishes accuracy benchmarks comparing design quality between specifications approved through EGM review and those developed with traditional agent workflows. The adoption story is strong internally, but quantified evidence of design quality improvement would strengthen the business case for large teams. Also monitor whether support expands to newer agent platforms. Finally, teams should test token consumption carefully before rolling out to production workflows, as the five-stage review cycle at scale could shift cost dynamics significantly.

For teams already embedding coding agents in their development cycle, testing EGM on a non-critical feature first makes practical sense. The design-first discipline and specification-level review offer real value if your harness supports isolated subagents and token budgets allow for the added review cost.

Related reading

• Letterpress-1: Coding agent for growth engineering
• dsh-superagent: Router-worker coding agent for parallel GitHub tasks
• TokenEconomizer cuts coding agent costs by 18-31%