How to Use GitHub Copilot Agent Mode

GitHub Copilot is the most widely adopted AI coding tool — roughly 15 million developers use it. In early 2026 it got a meaningful upgrade: Claude and Codex model access opened to all plan tiers (including Pro at $10/mo), and Copilot Workspace — the agent mode that works directly from issues and pull requests — matured into a real end-to-end workflow.

This guide is specifically about the agent side of Copilot: how to go from a GitHub issue to a working PR with Copilot doing the middle. If you just want autocomplete, you already have it.

What is Copilot Workspace / agent mode?

Copilot Workspace takes a GitHub issue as input, plans a solution, edits files, runs tests, and produces a PR. It lives at github.com/copilot and inside VS Code’s Copilot Chat. You still review every diff — you always review every diff — but the scaffolding is done for you.

Step 1 — Make sure your plan has agent mode

Agent mode is on Pro and above. Free tier gets autocomplete and chat but not the multi-file agent. Business/Enterprise adds audit logs, IP indemnity, and admin controls. Check your plan at github.com/settings/copilot.

Step 2 — Pick the right model

Since the February 2026 update, Copilot’s model picker exposes Claude and Codex options across plans. For agent-mode work, pick:

• Claude Opus 4.7 — deep reasoning, multi-file changes.
• GPT-5 / Codex — structured edits, good tool use.
• Claude Sonnet — faster, cheaper for routine changes.

You can switch mid-session. For a refactor, start with Opus/GPT-5; for “add a null check and a test,” Sonnet is fine.

Step 3 — Write the issue like a brief

The quality of the agent’s output is bounded by the quality of the issue. Good issues include:

• What’s broken or missing (with an example).
• What the expected behaviour is.
• Which files or areas to look at.
• Any constraints (don’t touch X, keep Y API stable).

Think of it like briefing a capable junior engineer you can’t talk to in real time. Vague issues produce vague PRs. Our prompt improver works on issue text too — paste the body, get a tighter version.

Step 4 — Hand the issue to Copilot

From a GitHub issue, click Assign to Copilot (or open it in Copilot Workspace). The agent produces a plan — the list of files it intends to change and why. Read this carefully before approving. 90% of bad PRs are bad plans that nobody questioned.

Step 5 — Review the plan, not just the diff

The plan is where you catch scope creep. If the agent says it’s about to touch 14 files for a 2-line bug fix, stop and redirect. Editing the plan is cheap. Editing the PR afterwards is expensive.

Step 6 — Let it run, then review like a human PR

Once you approve the plan, Copilot writes the code, runs tests (if configured), and pushes a branch. It opens the PR with an AI-generated description. Review it exactly like you’d review a teammate’s work:

• Read the actual diff, not just the summary.
• Run the code locally if it touches anything risky.
• Check tests actually test the change, not just pass.
• Look for unrelated edits you didn’t ask for.

Step 7 — Iterate by comment

Copilot responds to PR comments. Instead of manually fixing, leave a review comment like “this logic should handle null” and the agent pushes a follow-up commit. For small nits, still edit manually — faster than a full agent round-trip.

Step 8 — Configure repo-level rules

Add a .github/copilot-instructions.md file. Copilot reads it automatically on every agent task. Put your repo conventions there — framework version, test command, style rules. Same idea as Cursor’s .cursorrules and Claude Code’s CLAUDE.md.

When to use Copilot agent mode vs. alternatives

Copilot wins when your work already lives on GitHub — issue, PR, review, merge. The tight integration is the feature. Claude Code wins for long reasoning over local repos. Cursor wins for in-editor flow. Most teams end up with all three and use whichever fits the task.

Security and privacy

Business and Enterprise Copilot plans don’t train on your code and provide IP indemnity. On Pro, double-check the privacy settings at settings/copilot before pointing the agent at anything sensitive. Never put API keys, secrets, or customer data in prompts or issues — that goes for any coding agent.