Use GitHub Copilot features

This is the largest GH-300 domain. Know where each feature runs, how it receives context, its autonomy, and which plan or policy governs access.

Image source: GitHub Docs, GitHub Copilot CLI.

Surface	Best fit	Key distinction
Inline suggestions	Completing code while typing	Accept, partially accept, cycle, or dismiss suggestions
Copilot Chat	Questions, explanations, and scoped changes	Uses conversation and selected workspace context
Agent mode	Multi-step implementation in an IDE	Selects files, proposes edits, and runs approved tools
Copilot coding agent	Delegated GitHub task	Works independently and opens a pull request
Copilot code review	Review feedback	Suggests improvements; humans retain approval
Copilot CLI	Terminal-native assistance	Supports commands, scripts, files, and interactive sessions
Copilot on GitHub	Pull requests, summaries, Spaces, and repository tasks	Uses GitHub-hosted context

flowchart TD A[Small completion] --> B[Inline suggestion] C[Question] --> D[Chat] E[Coordinated local edits] --> F[Agent mode] G[Delegated issue] --> H[Coding agent pull request] I[Terminal workflow] --> J[Copilot CLI]

IDE context and customization

Context can include the active file, selection, open files, chat history, repository instructions, attached references, and tool results. Give only relevant context and state the goal, constraints, and acceptance criteria.

• Repository and path-specific instructions provide persistent conventions.
• Prompt files package repeatable task instructions.
• Custom agents define specialized roles, tools, and workflows.
• Model Context Protocol (MCP) servers expose approved external tools and data.
• Subagents isolate specialized work and conserve parent-agent context.

Instructions guide model behavior; permissions, policies, branch rules, and review processes enforce controls.

Ask, Plan, and Agent modes

Microsoft Learn distinguishes three built-in local roles. Choose by the result you need and the amount of change Copilot should be allowed to make.

Mode	What it does	Use it when
Ask	Explains and analyzes without changing files or invoking tools	Comparing approaches, finding edge cases, or understanding code
Plan	Researches with read-only tools and creates an implementation plan	Requirements or sequencing need review before edits begin
Agent	Edits files, invokes tools, runs commands, and iterates on failures	The goal is defined and implementation should proceed

Plan mode does not make code changes. After reviewing its questions and proposed steps, you can start implementation in Agent mode or open the plan as Markdown. Agent mode maintains a working set, streams edits, and can self-correct from compiler or test output, but its work still requires review.

Context controls in chat

• Use Add Context, drag and drop, or # mentions to attach files, folders, symbols, and selections.
• Use #selection for a precise selected block and #codebase when repository search is necessary.
• Use slash commands for repeatable operations; available commands depend on the client and current product version.
• Inspect the response's references to understand which files and instructions informed it.
• Start a new thread when prior conversation no longer supports the task.

CLI and administration

Follow the current Copilot CLI documentation for installation and authentication. In a session, inspect requested permissions, approve only necessary tool use, and validate changed files and generated scripts.

Copilot CLI supports two interfaces:

Interface	Invocation	Behavior
Interactive	Run copilot	Holds a steerable conversation with ask/execute and plan modes
Programmatic	Run copilot -p "PROMPT"	Performs one task and exits; useful in constrained automation

In interactive mode, use Shift+Tab to cycle modes. Local sandboxing can restrict filesystem, network, and system access; cloud sandboxing provides an isolated hosted session. Preview status and availability can change, so verify the current documentation.

Permission flags deserve special attention. --allow-tool grants a named tool without repeated approval, --deny-tool takes precedence and blocks a tool, and --allow-all-tools removes individual approval prompts. Broad automatic approval gives the agent the same effective command and file access as the user, so prefer narrow permissions and a sandbox for unattended work. Launch the CLI only inside a trusted, appropriately scoped directory.

Customization can include repository instructions, MCP servers, custom agents, hooks, skills, and Copilot Memory. Hooks can add deterministic validation or logging around agent actions; MCP extends available data and tools; memory can retain repository conventions. Each changes context or capability, not the developer's obligation to validate results.

Administrative need	Relevant control
Assign access	Subscription and seat management
Permit capabilities	Organization or enterprise Copilot policies
Investigate changes	Audit log events
Automate seat management	GitHub REST API for Copilot
Standardize reviews	Code review policy and custom instructions

Plans change frequently. Compare current GitHub Copilot plans rather than memorizing a stale feature matrix.

Readiness check

• When is agent mode a better fit than an inline suggestion?
• How does a coding agent normally return work to a team?
• Where can an administrator investigate policy activity?

Suggested answers

1. For multi-step work requiring coordinated edits, tools, and validation.
2. Through a pull request from its isolated environment.
3. In the audit log and relevant organization or enterprise settings.