GitHub Adds Rubber Duck Debugging AI to Copilot CLI for Smarter Code Reviews

GitHub is introducing a new experimental feature in GitHub Copilot CLI called Rubber Duck, designed to rethink how AI-assisted coding works. The feature builds on the classic rubber duck debugging concept, but replaces the duck with a second AI model that actively reviews decisions.

Instead of relying on a single model, Copilot now uses two. The primary LLM handles coding and planning, while a secondary model, called Rubber Duck, steps in to review key decisions and provide feedback at critical moments.

With the Rubber Duck method, one AI writes and another reviews

This approach targets a known issue with AI coding tools: early mistakes. When a model makes a flawed decision during planning, that error often compounds across dozens of steps. Since the same model evaluates its own work, it can miss its own blind spots.

Rubber Duck changes that dynamic by introducing a separate model from a different AI family. Acting as an external reviewer, it checks logic, flags potential issues, and helps prevent cascading errors.

Built for complex coding workflows

The feature shows the biggest impact in complex coding scenarios, especially multi-file projects or workflows that involve 70 or more steps. GitHub says the system reduced the performance gap between Claude Sonnet and Opus by 74.7%, highlighting its effectiveness in long, structured tasks.

Rubber Duck activates at key stages, including after planning, during complex implementations, and after test generation. It can also step in when the agent appears stuck in loops. Users can rely on automatic triggers or manually invoke them when needed.

To use the feature, developers must enable it through the /experimental command. The setup requires Claude as the primary agent, while GPT-5.4 powers the Rubber Duck reviewer.

In other news, GitHub has recently faced scrutiny over reports of promotional content appearing in pull requests. At the same time, threat actors have been distributing malware on the platform disguised as leaked Claude code.

Via Neowin