VS Code 1.136 adds Agent Merge to finish pull requests

Visual Studio Code 1.136 adds a preview feature called Agent Merge. It hands an open pull request to an AI agent, which then answers review comments, fixes failing checks and resolves merge conflicts until the request is ready. Microsoft dated the release notes September 2, 2026.

A pull request is a proposed code change waiting for review. Getting one merged often means a slow loop of comments, edits and re-run tests. Agent Merge aims at that loop, not at writing the code in the first place.

What Agent Merge actually does

Microsoft's release notes describe the feature in one line. "Agent Merge helps you take a pull request across the finish line," they say. "It asks an agent to address review feedback, fix failed checks and merge conflicts, and rerun workflows."

Four separate jobs sit inside that sentence.

Job	What the agent does
Review feedback	Reads reviewer comments and edits the code
Failed checks	Diagnoses a failing test or lint run and fixes it
Merge conflicts	Resolves conflicts against the target branch
Workflows	Re-runs the checks, then repeats the cycle

The agent does not merge anything by itself. Approval stays with a person. The release notes do not say which git hosting providers are supported.

How to turn it on

Agent Merge is off by default. Microsoft gates it behind a setting named chat.agentMerge.enabled.

Switching that setting on does not start anything either. You still enable the feature one session at a time. There are three ways in: the Agent Merge button, the command "Enable Agent Merge for Active Session", or the Agents window.

That per-session design is worth noticing. This is not a background service watching your repository and pushing commits.

The rest of the agent changes in 1.136

Agent Merge arrives alongside several smaller changes to how sessions work.

• Multi-root workspace support, marked experimental, for both Copilot and Claude agent sessions
• Workspace resolution that lets an agent identify a project by name
• A session hierarchy that shows related chats as children of a parent session
• Notifications when a session is waiting on your approval
• Readable breadcrumbs for files a session created
• A redesigned session input with the controls gathered in one place

The chat surface changed too. The release adds experimental chat backgrounds, dictation controls for enterprise administrators, and screen reader improvements.

What this means for developers

The interesting shift here is not code generation. It is that the agent now operates your continuous integration and your pull request queue. Those are shared systems, and the blast radius is wider than an editor buffer.

Watch the conflict resolution most closely. Of the four jobs, it is the one where a wrong answer still compiles and still passes tests. An agent can resolve a conflict by quietly dropping somebody else's change, and nothing downstream will complain. Read those diffs line by line, the way you would read a rebase you did not perform.

Budget is the second thing to check. The loop reruns workflows until checks pass, so a flaky test becomes an agent retrying against a test that fails at random. If your CI bills by the minute, put a cap somewhere before you turn this on. A synthetic worked example, published by Telemetry's creator on DuckDB's discussion board, puts cost per accepted task at $0.75 when failed attempts count and $0.30 when they do not. We covered how GitHub started billing Copilot code review on Azure Repos, and the same question applies to anything that reruns pipelines on your behalf.

Your real controls are unchanged, and they are not in this settings file. Branch protection rules, required reviewers and required status checks still decide what can land. The setting governs whether an agent may push more commits to the branch. It does not govern what your repository accepts.

Treat the preview label seriously for now. Enable it on a branch you own, on a pull request that is already close to green, and read every commit it produces. Research on coding agents found they disagree about which tool to reach for far more often than their marketing suggests. An agent that is confident and wrong about a merge conflict is the failure mode to plan around.