Claude Code: Background Agents Now Auto-Commit, Push, and Open Draft PRs
Claude CodeBackground agents launched from claude agents no longer stop to ask what to do once code work is finished in a worktree: they now automatically commit the changes, push the branch, and open a draft pull request. This removes a manual handoff step that previously required the developer to return to the session and approve next steps before the work could land anywhere reviewable. Opening the PR as a draft avoids triggering CI pipelines or notifications configured for standard pull requests, giving developers a safe, low-noise way to review autonomous agent output.
Key Takeaways
- Background agents now auto-commit, push, and open a draft PR the moment code work finishes in a worktree, instead of pausing for manual confirmation.
- Draft PR status is intentional, preventing autonomous agent runs from triggering CI, reviewer notifications, or other automation tied to non-draft PRs.
- This removes a manual checkpoint that previously required returning to the session to decide what to do with completed work.
- The change pairs with new background-agent notifications, which fire when a session needs input or completes, closing the loop on unattended monitoring.
- This is part of a broader v2.1.198 push toward hands-off multi-agent workflows, alongside changes to how subagents inherit model and reasoning configuration.
Sources & Mentions
1 external resource covering this update
Autonomous Agents Now Finish the Job
Previously, when a Claude Code background agent completed code changes inside a worktree, it would stop and wait for the user to decide what to do next: commit, push, open a PR, or discard the work. As of v2.1.198, that manual checkpoint is gone for the common case: background agents now commit their changes, push the branch, and open a draft pull request automatically once code work is finished.
Why Draft, Not Ready-for-Review
Opening the PR in draft state is a deliberate choice. Many teams have automation tied to non-draft pull requests: CI runs, reviewer pings, Slack notifications, and having every autonomous agent run trigger those workflows would create significant noise. By defaulting to draft, Claude Code lets the agent's output land somewhere reviewable and diffable without immediately pulling in reviewers or kicking off downstream automation the developer hasn't opted into yet.
Part of a Broader Push on Unattended Work
This change sits alongside other v2.1.198 improvements to the claude agents experience, including notifications when an agent needs input or finishes and improvements to how the Explore agent and subagents share reasoning configuration with the parent session. Together, these changes reflect a broader effort to make background and multi-agent workflows genuinely hands-off, rather than requiring the developer to babysit the session until it can be handed off manually.