At a glance
GitHub's terminal coding agent supports interactive and programmatic execution, granular permissions, local and cloud sandboxes, MCP, custom agents, parallel subagents, file rollback, and GitHub-hosted or custom local models.
Good choice if
- GitHub-centered development with repository, issue, pull-request, and MCP context
- Parallel implementation plans using fleet subagents and separate context windows
- Teams that need programmatic runs, auditable permissions, rollback, and optional isolation
Check before choosing
- A Copilot plan or compatible custom provider configuration is required
- Local and cloud sandbox features are in public preview, opt-in, and have documented platform and lifecycle limits
- As of 1.0.77, unconditional approval in autopilot disables the sandbox for the current session when bypass is allowed; administrators can separately enforce sandbox use through native MDM policy
See 6 more considerations
- Since 1.0.78, `/rewind` restores only Copilot-written files whose contents still match its last write, so later manual edits are skipped; shell, Git, and external side effects remain outside recovery
- Git worktrees isolate repository files rather than processes, credentials, or network access; since 1.0.79, `/worktree`, `/worktree new`, and `--worktree` default to the current HEAD unless `worktreeBaseRef` selects the remote default branch
- Server-managed settings fall back to a persistent cache on fetch failure and start without the unconfirmed restriction when no usable cache exists, unless `forceRemoteSettingsRefresh` is enabled
- Copilot Memory is a hosted public-preview feature with a 28-day unused-item retention policy rather than local durable project state
- Fleet increases GitHub AI Credit use because each subagent calls a model independently
- The public repository is a support and binary-distribution surface, not the proprietary agent implementation or product evaluation suite