Onboarding your AI peer programmer: Setting up GitHub Copilot coding agent for success

A post on GitHub’s blog, published on July 31, 2025 by a GitHub senior developer advocate, on how to prepare a repository for GitHub Copilot coding agent. Its organizing analogy is that much of the setup resembles onboarding a new developer — good documentation and a streamlined setup process — while a few things are specific to Copilot being an AI.

The post first distinguishes Copilot’s two agentic capabilities: coding agent, which it describes as an autonomous tool that takes an issue, spins up an Actions container, iterates in the background and returns a pull request, and agent mode, an interactive sidekick in the editor or on github.com that executes smaller multi-step tasks with the developer in real time. It then follows the coding agent’s workflow step by step and says, for each step, what a team can do to help it produce a better pull request.

Key Points

• When assigned an issue, the coding agent follows a set pattern: it creates a branch, creates a pull request to track its work, creates a contained environment running inside GitHub Actions, reads the issue or prompt, explores the project, works iteratively toward a solution, and finally updates the pull request and notifies the team it is ready for review.
• The agent’s environment is configured with a workflow file at .github/workflows/copilot-setup-steps.yml containing a job that must be named copilot-setup-steps, listing the steps that install whatever the environment needs; the post’s example sets up Python dependencies and SQLite. The author suggests reusing an existing workflow that already creates a development environment.
• The author advises telling Copilot how something should be done rather than leaving it to install requisite services on its own, which may lead to unexpected versions or other mistakes.
• The more clearly defined the issue, the better the pull request; the post recommends including a clear problem statement or user story, full error output and reproduction steps for bugs, relevant history, and suggestions on how to approach the issue, and gives a sample issue for migrating tests from unittest to pytest.
• Standard documentation and project-structure practice helps the agent the same way it helps developers: when it takes an issue it first explores the codebase, reading README files and searching for related code, so an up-to-date README, code comments, good naming and a logical folder structure give it a more predictable environment.
• Custom instructions come in two kinds: a repository-wide .github/copilot-instructions.md applied to all requests, and <file-name>.instructions.md files under .github/instructions/ whose applyTo glob targets specific files. The post suggests the repository-wide file cover what is being built, user stories, frameworks and libraries, project structure and global coding guidelines, and recommends targeted instructions files for correcting particular kinds of mistake the agent keeps making.
• The agent can use Model Context Protocol servers; the GitHub and Playwright MCP servers are available by default, an existing .vscode/mcp.json in the project can be used to identify MCP servers, and otherwise servers are configured in the repository’s Copilot coding agent settings. The post’s example enables the Azure MCP server restricted to Bicep schema support.
• The agent has a default firewall limiting its internet access to core services such as the npm and pip package hosts, which the post presents as a way to manage data exfiltration risk should malicious instructions reach the agent; adding a remote MCP server or other internet access means updating the firewall’s allow list.
• The pull request’s “View session” button shows everything the agent did, which the author recommends both for validating its work and for refining how tasks are assigned and its environment configured.

Context

This is guidance from a GitHub developer advocate about GitHub’s own product, drawn from the author’s experience rather than measured results. The author notes that investing in instructions files also helps when Copilot is used in the IDE, not only when tasks are assigned to the coding agent.