Rubber Duck

Rubber Duck is an experimental feature for the GitHub Copilot CLI that introduces a second‑model reviewer into the coding‑agent workflow. When a primary model such as Claude Sonnet generates a plan or code, Rubber Duck—drawn from a different AI family (for example GPT‑5.4 when Claude is the orchestrator)—independently assesses the output at key decision points, aiming to catch errors that the original model’s training biases might miss.

The tool targets developers who use Copilot CLI for complex, multi‑file or long‑running tasks and want an additional safety net before code is executed. By providing a cross‑family perspective, Rubber Duck helps identify planning inefficiencies and confident mistakes early, reducing the need for extensive later‑stage fixes.

Rubber Duck is accessed via the /experimental flag in Copilot CLI and operates alongside other experimental capabilities. Early evaluations show that pairing Claude Sonnet with Rubber Duck narrows the performance gap to Opus by roughly three‑quarters, indicating measurable improvement on difficult tasks.