{"slug": "vs-code-1-135-rubber-duck-agent-host-protocol-sessions", "title": "VS Code 1.135: Rubber Duck, Agent Host Protocol, Sessions", "summary": "Microsoft released VS Code 1.135 on August 26, introducing the open, MIT-licensed Agent Host Protocol (AHP) that turns the editor into a universal agent runtime, external session continuity for picking up Claude Code or Copilot work, and Rubber Duck, an experimental second-model reviewer that closed 74.7% of the performance gap between Claude Sonnet and Opus on SWE-Bench Pro. The release marks VS Code's commitment to an agent-first architecture, with AHP positioned as an LSP-like standard for AI agents.", "body_md": "VS Code 1.135 dropped August 26 with three features that belong together: an open Agent Host Protocol that turns VS Code into a universal agent runtime, external session continuity that lets you pick up Claude Code or Copilot work without starting over, and Rubber Duck — an experimental second-model reviewer that checks the primary agent’s output before it ships. This is not a maintenance release. It is VS Code committing to an agent-first architecture, and the open standard at the center of it is the part that will matter most six months from now.\n\n## Agent Host Protocol: VS Code’s LSP Moment for AI Agents\n\nThe most significant thing in this release is not user-facing. The Agent Host Protocol (AHP) is an open, MIT-licensed standard that Microsoft published alongside VS Code’s own implementation. It positions next to the [Language Server Protocol](https://microsoft.github.io/language-server-protocol/) and the Debug Adapter Protocol as shared infrastructure — the kind of glue that the whole ecosystem can build on.\n\nAHP uses JSON-RPC for communication and immutable state with pure reducers for session synchronization. Sessions are URI-addressed and persistent: they survive the editor closing, they can run on a remote machine, and multiple clients can observe and control the same session simultaneously. Agents run in a dedicated process, separate from extensions, so a busy agent does not block the editor UI.\n\nThe spec is [published on GitHub under MIT license](https://microsoft.github.io/agent-host-protocol/). Other editors — JetBrains, Zed, Neovim — can implement either side. VS Code is betting that an open standard is more powerful than a proprietary moat. That is the same bet Microsoft made with LSP in 2016, and it won. Ecosystem adoption of AHP has already started: an AHP plugin for opencode appeared within days of the release. The VS Code team’s [architecture deep-dive](https://code.visualstudio.com/blogs/2026/08/26/agent-host-architecture) explains the full design.\n\n## External Sessions: Continue Anywhere\n\nThe first user-facing feature enabled by AHP is external session continuity. If you started a debugging or refactoring session in Claude Code, Copilot CLI, the GitHub Copilot app, or Codex, VS Code 1.135 can pick it up. You find recent sessions under **Copilot > Agents > Recent Sessions** and continue from exactly where the other app left off — no context loss, no re-explaining.\n\nThis addresses a real friction point. Developers switch tools mid-task constantly, and until now, crossing the boundary between a CLI agent and an IDE meant starting the conversation over. The practical upside is immediate; the longer-term implication is that VS Code is positioning itself as the canonical place to manage agent work, regardless of where it started.\n\n## Rubber Duck: A Second Model That Argues Back\n\nRubber Duck is the feature with the most immediate quality impact. Run `/rubber-duck`\n\nin any Copilot agent host session, and a complementary model — not the same one driving your session — reviews the agent’s plans, code, and tests. When Claude is the orchestrator, GPT-5.4 runs the review. The result is cross-provider critique baked into your workflow.\n\nRubber Duck is read-only: it reports, it does not modify. Feedback comes in three tiers: blocking issues (logic errors, security vulnerabilities), non-blocking issues (design concerns, maintainability), and suggestions. The benchmark data is worth taking seriously: on [SWE-Bench Pro](https://www.swebench.com/), Claude Sonnet paired with Rubber Duck closed 74.7% of the performance gap between Sonnet and Opus. That is a meaningful quality lift for the cost of running an extra review pass.\n\nThis feature is marked Experimental. The model pairings and command interface may change. It works now and is worth trying, but do not build automated pipelines around it until it stabilizes.\n\n## Token Usage: Finally, Full Visibility\n\nThe chat response footer now shows something useful when you hover: a per-model breakdown of input tokens, cached input tokens, and output tokens for each turn. The context window indicator shows how much of the model’s context window the session is consuming, with a breakdown by category and total AI credits on hover.\n\nThis is table stakes at this point. As developers run longer agent sessions and pay per token, knowing exactly what you are spending and where is not optional. The cached input line is particularly useful — it shows whether your prompts are actually hitting the cache or burning fresh tokens every time.\n\n## What It All Adds Up To\n\nVS Code 1.135 is coherent in a way that monthly releases often are not. The Agent Host Protocol is the foundation, external sessions are the first application of it, Rubber Duck is the quality layer, and token transparency is the cost control layer. These features are not independent checkboxes — they are a stack.\n\nThe open standard is the piece to watch. If AHP follows the LSP trajectory, the conversation about which editor or agent tool you use will gradually matter less, because sessions will move freely between them. VS Code is making that bet publicly and early. Whether competing editors adopt it or build their own protocol in response, the market for interoperable agent tooling is moving.\n\nUpdate to 1.135 via [the official release notes](https://code.visualstudio.com/updates/v1_135), enable the Agent Host in settings, and run `/rubber-duck`\n\non your next complex agent session. The quality data suggests it will catch something.", "url": "https://wpnews.pro/news/vs-code-1-135-rubber-duck-agent-host-protocol-sessions", "canonical_source": "https://byteiota.com/vs-code-1-135-rubber-duck-agent-host-protocol/", "published_at": "2026-08-29 18:09:51+00:00", "updated_at": "2026-08-29 18:18:39.721461+00:00", "lang": "en", "topics": ["developer-tools", "ai-agents", "ai-infrastructure"], "entities": ["Microsoft", "VS Code", "Agent Host Protocol", "Rubber Duck", "Claude Code", "GitHub Copilot", "SWE-Bench Pro"], "alternates": {"html": "https://wpnews.pro/news/vs-code-1-135-rubber-duck-agent-host-protocol-sessions", "markdown": "https://wpnews.pro/news/vs-code-1-135-rubber-duck-agent-host-protocol-sessions.md", "text": "https://wpnews.pro/news/vs-code-1-135-rubber-duck-agent-host-protocol-sessions.txt", "jsonld": "https://wpnews.pro/news/vs-code-1-135-rubber-duck-agent-host-protocol-sessions.jsonld"}}