Claude Code v2.1.139 shipped a /goal command that tells the tool to keep working until a task is done - no mid-task check-ins required.
The previous behavior required staying present: Claude Code would complete one action and stop, waiting for you to review and prompt again. For small edits that's fine. For anything spanning multiple files or needing a dozen steps to complete, it turns into a babysitting session. /goal removes that friction by running the full sequence autonomously, only pausing if it genuinely gets stuck or encounters an ambiguity it can't resolve on its own.
What autonomous mode changes in practice
The /goal command treats your instruction as a destination rather than a single step. Refactoring a module, writing a test suite for a new feature, or migrating a configuration pattern across a codebase - these are all tasks with a clear end state but many intermediate steps. With the old flow, each handoff was manual. With /goal, Claude Code chains them together and reports back when it's finished or needs a decision.
This works best on scoped, verifiable work: a new file, a batch of tests, a migration script. It's less suited to exploratory sessions where you're working out the direction as you go.
The real cost of running unsupervised
Autonomous coding modes amplify both speed and errors. If Claude Code misreads a task and goes the wrong direction, the further it runs before you notice, the more there is to untangle. Twenty cascading changes built on one wrong assumption can take longer to fix than the original task.
For anything touching production logic or making irreversible changes, the old step-by-step flow is still the safer default. Save /goal for work where the output is easy to audit at the end.
To use it: upgrade to v2.1.139 (claude --version to confirm), then run /goal [your task description] to start a session.