GitHub Copilot's Autopilot Mode Lets Agents Run to Completion Without Asking Permission

GitHub published the March VS Code releases for Copilot on April 8, covering versions 1.111 through 1.115. The standout feature is Autopilot, now in public preview: a permission mode where the agent handles decision-making itself and runs until the task is complete.

Three Permission Levels

The update introduces a permissions picker in Copilot Chat that controls how much autonomy the agent has during a session.

Default keeps the existing behavior: the agent asks for approval before taking actions and waits for answers to questions.

Bypass Approvals skips the tool confirmation dialogs but still stops for user questions. Think of it as silencing the “are you sure?” prompts without going fully hands-off.

Autopilot removes both. The agent approves its own actions, handles ambiguity on its own, and retries automatically when it hits an error. It runs in a loop until it decides the task is done by calling a task_complete action. You kick it off with a single prompt and come back to the result.

This is the same pattern that Claude Code uses with --dangerously-skip-permissions, Codex uses with its auto-approve mode, and Cursor has shipped in its own agent configurations. GitHub Copilot now has a named, in-UI version of the same idea.

How Autopilot Handles Errors

The agent doesn’t just stop when something goes wrong. If a dependency is missing, it tries to add it. If a command fails, it reads the error and adjusts. The goal is to stay in the loop until the task is actually finished rather than surfacing every obstacle back to the user.

That’s the behavior change worth paying attention to. Previous agent modes in Copilot still required human intervention at each decision point. Autopilot is designed to handle those points itself, for better or worse.

Configurable Reasoning Depth

The March update also adds a reasoning effort control in the model picker. For models that support extended thinking, such as Claude Sonnet 4.6 and GPT-5.4, you can choose how deeply the model reasons before responding. Higher reasoning effort takes longer but tends to produce better results on complex tasks. Lower effort is faster and good enough for routine work.

This is available directly from the model picker inside the chat panel, no settings file required.

Nested Subagents

Agents can now invoke other agents as subagents, which enables more complex workflows where a coordinating agent delegates specialized subtasks. The practical limit is how much context and compute you’re willing to spend, but the capability is there.

Integrated Browser Debugging

A new browser debugging integration lets you set breakpoints and inspect variables in a running web application without leaving VS Code. This is separate from the Autopilot feature but ships in the same release. If you’re doing full-stack work where Claude or Copilot is writing code that runs in a browser, you can now catch runtime issues inside the same editor where the code was written.

Other March Additions

The full changelog covers a lot of ground:

Image and video support in chat: You can now drop screenshots, recordings, or design assets directly into the chat panel for context
Unified customizations editor (preview): One place to manage all instructions, custom agents, skills, and plugins, with a built-in MCP and plugin marketplace browser
Session forking: Start an alternative branch from any point in a conversation to explore different approaches without losing the original thread
Faster #codebase search: Now runs purely semantic searches against a single auto-managed index rather than the older hybrid approach
Agent-scoped hooks (preview): Pre- and post-processing logic that fires around agent actions, useful for logging, validation, or side effects
TypeScript 6.0 support: Updated language services for projects on the new TypeScript release

Context

GitHub Copilot’s Autopilot mode lands at a time when fully autonomous coding agents have become a practical feature category rather than a research concept. Claude Code’s routines, Codex’s cloud execution, and Cursor 3’s agents window all ship variations on the same theme: start a task, let the agent run, review the result.

Copilot’s version is more conservative by default (the permission picker defaults to the existing approval workflow) but Autopilot mode is there for users who want it. Given that Copilot has 29% workplace adoption across professional developers, this feature lands in front of a large audience quickly.