Copilot CLI Becomes an Agent: Writes Code, Runs Tests, and Fixes Bugs Without Prompting

GitHub Copilot CLI entered general availability for all Copilot subscribers in February 2026 — and it arrived differently than when it entered preview in September 2025.

It is no longer just a terminal assistant. It is an autonomous development agent that plans, executes, tests, reviews, and iterates — without leaving the command line.

What is Autopilot mode

The main novelty of the GA is the Autopilot mode: after providing the initial instruction, Copilot CLI executes all necessary steps without asking for approval at each stage. It runs commands, evaluates outputs, fixes intermediate errors, and iterates until the goal is complete.

To activate, simply press Shift+Tab during an interactive session. The mode is especially useful for well-defined tasks: writing tests for an entire module, refactoring files, fixing CI failures, or performing batch operations in scripts.

Specialized agents and automatic delegation

The CLI now automatically delegates to specialized agents according to the task type:

— Explore: rapid codebase analysis and specific queries.

— Task: execution of builds and tests.

— Code Review: high-signal change review.

— Plan: implementation planning before execution.

Multiple agents can run in parallel. Prefixing a prompt with & delegates to the coding agent in the cloud, freeing up the local terminal for other tasks.

Cross-session memory

One of the most relevant additions is persistent memory. Copilot CLI now remembers conventions, patterns, and preferences it has learned about the codebase across different sessions — without needing to recontextualize from scratch every time.

Cross-session memory also allows querying previous work, files, and pull requests over time.

MCP and available models

The CLI comes with the GitHub MCP server integrated by default, and supports custom MCP servers to connect any external tool. Plugins can be installed directly from GitHub repositories with /plugin install owner/repo.

In terms of models, the CLI allows choosing between Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.3-Codex, and Gemini 3 Pro — with premium request cost multipliers varying by model.

What this means in practice

The Copilot CLI GA is another step toward workflows where the dev defines the goal and the agent executes. It is not simple task automation — it is autonomous iteration on complex tasks, with intermediate error self-correction.

The terminal — for decades the most powerful and least user-friendly development interface — is starting to become an environment where AI bridges the gap between intent and execution.

Available to all paid GitHub Copilot subscribers via npm, Homebrew, or WinGet.