Brainstorm with the Rubber Duck agent in GitHub Copilot CLI

If you’ve used coding agents for anything non-trivial, you’ve probably done this: let one model (say Claude Sonnet 4.6) do the work, then ask a different model (say GPT 5.4) to review it. Different model families have different blind spots, so a cross-family review catches things self-reflection won’t.

GitHub Copilot CLI now has this built in 🎉

Rubber Duck is a review agent that uses a model from a different family than your primary session. Right now, it pairs Claude orchestrators with GPT-5.4 as the reviewer, with more combinations on the way. Its job is to check the agent’s work and surface concerns: missed details, questionable assumptions, edge cases.

Here’s when it kicks in:

1️⃣ After drafting a plan - Catching a bad plan early saves you from digging a deeper hole. This is probably where you’ll see the biggest wins.

2️⃣ After a complex implementation - A second set of eyes on tricky code to catch edge cases.

3️⃣ After writing tests, before running them - So you don’t get that false comfort of green checkmarks when the tests themselves are wrong.

It also steps in if the agent starts going in circles. And you can trigger it manually anytime by just asking Copilot to critique its work.

In the short video below, you can see Rubber Duck pop up while I was building an agent-based sample app with Azure Cosmos DB. I didn’t ask for it. It decided to review the code on its own before wrapping up 🦆

⚡ To try it: /experimental in GitHub Copilot CLI

🔗 Link to the blog post - https://github.blog/ai-and-ml/github-copilot/github-copilot-cli-combines-model-families-for-a-second-opinion