VS Code 1.135 Brings Rubber Duck Debugging to AI Agents

VS Code Gets a Second Opinion

Microsoft released Visual Studio Code 1.135 on 26 August 2026. The headline feature is an experimental mode called Rubber Duck — named after the classic debugging technique of explaining your code to an inanimate object, which forces precision and often surfaces the error. VS Code 1.135 makes that idea computational: a secondary AI model, different from the one driving the primary coding agent, reads the agent's output and surfaces what it might have missed.

What Rubber Duck Actually Does

When a developer is working with an AI coding agent — GitHub Copilot, Claude, or another session — and generates a response, Rubber Duck can be invoked to run a review pass using a complementary model. The complementary model has not seen the same prompt context that drove the primary agent's response, which means it approaches the output fresh. Its task is to flag missed edge cases, overlooked assumptions, gaps in test coverage, or code paths where the primary agent's solution fails to account for realistic inputs.

The design premise is that two models with different training characteristics are unlikely to share the same blind spots. A second model trained differently will catch what the first routinely misses. Rubber Duck makes that second pass a one-click operation within the editor without requiring the developer to open a separate tool or re-enter context.

Why This Matters for Production Code

Agentic coding workflows have accelerated sharply in 2026, and so has the volume of AI-generated code reaching production without sufficient review. The failure mode is not obviously incorrect code — AI agents generally produce syntactically sound, test-passing output. The problem is plausible code that fails under edge conditions the agent did not consider. Rubber Duck is a systematic response: not a replacement for human review, but a low-friction first filter that catches a class of errors before they reach code reviewers or CI pipelines.

The Redesigned Agents Window

VS Code 1.135 also ships a redesigned Agents window. The previous layout spread agent controls and session information across multiple panes, creating friction when switching between sessions or reviewing session state. The 1.135 release consolidates agent controls into a single-pane layout with simplified session controls and session information that is easier to locate at a glance. The redesign reflects the broader trajectory of VS Code's Copilot integration: the editor is increasingly being optimised for multi-agent workflows where developers may run several parallel agent sessions simultaneously.

External Agent Sessions

A third addition in 1.135 is the ability to continue recent Copilot or Claude agent sessions from other applications inside VS Code. A developer who started a problem-solving session in the Claude desktop app or a browser-based Copilot chat can pick it up in the editor without losing context. This matters because agentic development increasingly spans surfaces — a developer might describe a problem in a chat interface, generate an initial plan, then want to execute and iterate in their editor with the full conversation context visible. VS Code 1.135 removes the friction of copying and re-entering that context manually.

Token-Level Usage Tracking

VS Code 1.135 adds a per-model breakdown of input, cached input, and output tokens for each chat turn. Developers can see, for each message sent to an AI agent, exactly how many tokens were consumed, how many were served from the prompt cache, and how many were generated. This is useful both for cost estimation — teams running Copilot or Claude at scale can identify token-heavy workflows and optimise them — and for debugging prompt efficiency. Unexpectedly high token counts often point to context management problems, such as unnecessarily long system prompts or inefficient retrieval patterns, that are worth addressing.

What This Means for Software Teams in India

For Indian software development teams, VS Code 1.135 has two particularly relevant additions. Rubber Duck is directly applicable to quality assurance in teams relying heavily on AI-generated code: an automated second-opinion pass reduces the load on human reviewers without requiring them to check every agent output individually. For teams operating under strict code review requirements — common in financial services, healthcare, and enterprise software contexts prevalent in the Indian IT industry — this automated first filter meaningfully reduces reviewer workload. The per-turn token tracking gives teams the visibility to manage AI usage costs as they scale from individual developer licences to team-wide agentic workflows with significant token consumption.

The Bottom Line

VS Code 1.135, released 26 August 2026, adds three substantive capabilities to AI-assisted development. Rubber Duck is an experimental second-opinion mode that runs a complementary AI model over the primary agent's output to surface missed edge cases — a practical response to the quality risks of agentic code generation at scale. The redesigned Agents window consolidates controls into a single pane. External agent session continuity lets developers carry context from Copilot or Claude sessions in other applications directly into VS Code. Per-turn token tracking gives teams cost visibility as agentic workflows scale. For Indian software teams investing in agentic development, 1.135 makes the tooling noticeably more production-ready.