VS Code 1.135 shipped August 26 with a feature that sounds like plumbing but is actually a structural change to how AI coding sessions work. The Agent Host Protocol (AHP) decouples agent sessions from the editor window — the same architectural pattern that made the Language Server Protocol the lingua franca of code intelligence in 2016. If LSP taught editors to share language servers, AHP teaches them to share AI agents. Your session is no longer owned by the window it started in.
What the Agent Host Protocol Actually Does #
The AHP is an open, MIT-licensed specification published by Microsoft on GitHub. It defines how a host — the process running your AI agent — communicates with one or more client windows over JSON-RPC and WebSocket. Sessions run in their own process, persist independently of any editor window, and use immutable state with sequence numbers to stay synchronized across multiple clients viewing the same session.
Microsoft explicitly frames AHP alongside the Language Server Protocol and Debug Adapter Protocol as the same kind of move: decouple a capability that used to belong to the editor and turn it into a protocol any tool can implement. LSP did this for language intelligence. DAP did it for debuggers. AHP aims to do it for agent sessions.
Pre-1.0 Warning: AHP is a draft protocol. Microsoft’s own documentation warns that wire types, actions, and state shapes may break before 1.0. It already has a third-party implementation from the opencode project, but do not bake AHP endpoints into production CI/CD pipelines until the specification stabilizes.
Portable Sessions: The Part You’ll Actually Use Today #
The most immediately useful feature in 1.135 is portable agent sessions. VS Code’s Sessions list can now display recent Copilot and Claude sessions that you started in other applications. Supported sources include Copilot CLI, the GitHub Copilot app, Claude Code, and Codex.
In practice: you start a refactor task in Claude Code in your terminal, get halfway through the planning conversation, then switch to VS Code to work with the visual file diff and editor. Instead of rebuilding context from scratch, you continue the same session. The conversation history, context window, and agent state carry over.
External sessions are hidden by default and require a Copilot subscription. You configure visibility in the Sessions panel: None, Recent (two most recent from the last seven days), Last 24 Hours, Last 7 Days, or All. Multiple VS Code windows can also connect to the same local session simultaneously — useful for pair programming with a shared agent context.
Rubber Duck: AI Reviewing AI, By Design #
Rubber Duck is a new experimental review agent that runs on a model from a different AI family than your primary orchestrator. If you’re running Claude Sonnet as your orchestrator, Rubber Duck runs on GPT-5.4. The reasoning: different model families are trained differently, so they miss different things. A complementary reviewer surfaces errors the primary model is likely to repeat.
Rubber Duck activates at three checkpoints: after drafting a plan, after complex multi-file implementations, and after writing tests but before running them. On SWE-Bench Pro benchmarks, Claude Sonnet paired with Rubber Duck closed 74.7% of the performance gap between Sonnet and Opus on the hardest problem class — tasks spanning three or more files requiring 70 or more steps.
That number deserves context: this is on difficult problems, not routine autocomplete. The benefit on average tasks is smaller. But if you regularly tackle large refactors or complex multi-service changes, the 4.8% lift on the hardest problems is real. Rubber Duck is off by default; enable it in the agent settings panel.
Token Usage Breakdown — Finally #
1.135 adds per-model token breakdowns for every chat turn. Hover over a response in the Chat view and you’ll see a split of input, cached input, and output tokens per model used in that turn. Hover over the context window indicator to get the cumulative session total in credits.
This matters most when Rubber Duck is active — you’re burning tokens on two models simultaneously. The breakdown lets you see exactly where credits are going and decide whether the Rubber Duck review is worth the overhead for a given task.
What This Means for the Editor Wars #
The editor fork market — Cursor, Antigravity, Kiro, and others — built much of its value proposition around owning the AI session natively. AHP changes the underlying premise: if sessions are portable and agent-agnostic, the editor becomes a client to the session rather than the owner of it. That’s a less defensible position for forks to hold. VS Code continues to borrow its moat back, one protocol at a time.
The full VS Code 1.135 release notes and the agent host architecture documentation cover the protocol details and adapter model for Copilot, Claude, and Codex. AHP 1.0 has no official ship date — the pre-1.0 period is the time to experiment, not to ship.