Discussion

Google’s Elephant-Goldfish plugin sends fresh AI agents to review the spec before the code

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#11914

Google has open-sourced a plugin that asks AI coding agents to write and review a specification before they write code. Its distinctive move is to send fresh, context-free agents to challenge the design, aiming to catch mistaken assumptions before they reach a pull request.

Watch Desk analysis

What happened

The Elephant-Goldfish Model plugin is a set of skills and review tools, not a new language model: developers supply the model, and the plugin uses their existing model quota. MarkTechPost describes a workflow that starts with brainstorming, develops a design document, then runs five sequential review stages, from a high-level critique to checks on implementation details.

The report says the plugin can work with several coding-agent harnesses, including Antigravity, Claude Code, Codex and ChatGPT. It also says the project is not an officially supported Google product. No accuracy benchmark is reported, so its test results show that the workflow ran, not that it reliably produces better software.

Why it matters

AI coding tools can move quickly from an idea to code, but a confident agent can also carry a bad assumption through the whole job. Reviewing the specification first gives developers a chance to spot gaps before implementation makes them more expensive to fix. Fresh reviewers are meant to bring less conversational baggage to that check, though the method depends on the host actually keeping their context separate.

That makes this a practical experiment in how to supervise coding agents, rather than another model launch. The trade-off is familiar: more review may catch more problems, but multiple agent runs can use substantial tokens and add time.

Our read

The clever part is moving scrutiny upstream, where a flawed design is still a document rather than a pull request with momentum. It is worth trying on projects where a missed requirement would be costly, but the reported test runs are not a substitute for measuring whether the reviews find important defects in real work. And if the agents share instructions they were meant to be isolated from, the goldfish may remember more than advertised.

What to watch

  • Whether the project publishes benchmark results comparing reviewed and unreviewed specifications.
  • Whether fresh-agent isolation holds across supported coding tools.
  • How much extra model usage and review time the workflow adds in everyday projects.

Discussion spark: Would you trust fresh AI agents to review a coding plan before implementation, or should a human remain the only meaningful design reviewer?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.