Visual Studio Code 1.135 introduces two notable changes for AI‑assisted development: an experimental “Rubber Duck” feature that invokes a second AI model to review the primary Copilot session, and an Agent Host Protocol that runs agent processes in a separate host, enabling persistent, cross‑window sessions. Both changes aim to give engineers more confidence in generated code and better operational insight.
Rubber Duck: a second‑model reviewer
The new Rubber Duck command (triggered with /rubber-duck) launches a complementary model inside the Copilot agent host. The secondary model examines the primary model’s plan, code, and test output, explicitly looking for gaps the first model may have missed. Microsoft frames this as a “second opinion” to surface edge cases before they reach production. For AI engineers and DevOps teams, the practical effect is an additional automated review step that can catch simple oversights without requiring a human reviewer at every iteration.
Agent Host Protocol and session continuity
VS Code now ships with an Agent Host Protocol (AHP) that isolates agent harnesses in a dedicated process. This design allows multiple VS Code windows to attach to the same agent session, preserving state across devices or monitors. The Sessions list also surfaces agent sessions that originated outside VS Code—such as Copilot or Claude sessions started in other tools—provided the chat.agentSessions.showExternal setting permits it. The implication for platform engineers is a more durable execution environment for long‑running AI tasks, reducing the need to restart agents when switching contexts.
Token‑usage visibility and UI refinements
Hovering over a chat response footer now displays a per‑model breakdown of input, cached input, and output tokens for that turn. This granular accounting helps finance or SRE teams track AI spend, especially when multiple models are in play. The Agents window has been simplified to a single‑pane layout, with session details (changes, PRs, issues, artifacts) presented as interactive pills above the chat input, and a cleaner diff view. These UI tweaks do not add new capabilities but reduce friction when monitoring agent activity.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
From an implementation perspective, the Rubber Duck feature should be treated as an optional, experimental safety net rather than a replacement for formal code review. Teams may want to benchmark the secondary model’s false‑positive rate and decide whether its suggestions merit automated gating. The Agent Host Protocol’s separate process model introduces a new operational surface: monitoring the host process, ensuring it restarts cleanly, and configuring the chat.agentSessions.showExternal flag to balance visibility with noise. Security engineers should note that the token‑usage overlay reveals model‑level consumption, which could be useful for detecting anomalous usage patterns but does not itself enforce any access control. Finally, the openness of AHP to third‑party agents remains an open question; organizations should evaluate vendor lock‑in risk if they plan to run non‑Microsoft agents alongside Copilot.
