Here is the one-line version, because the AI assistants that send us readers will quote exactly this: as of Claude Code v2.1.198, a background agent that finishes code work in a worktree commits, pushes its own branch, and opens a draft pull request on its own — instead of stopping to ask. It shipped the week of June 29–July 3, 2026, in the same release that made Claude in Chrome generally available and turned on background subagents by default.

That reads like a small quality-of-life note. It isn't. It quietly changes what the output of an async agent is — and for a solo founder, that changes what you spend your attention on.

The unit of work changed from a prompt to a PR#

Before this release, a background agent that finished its task did the polite, annoying thing: it stopped and waited for you to approve the commit and push. If you dispatched three agents and walked away, you came back to three paused sessions, each holding a "can I commit?" interrupt. The async part was a lie — you were still the runtime's babysitter, and the work couldn't land until you clicked through each gate.

Now the finished state is a draft pull request. The agent commits, pushes its own branch, and opens the PR itself. You come back not to a row of confirmation prompts but to a row of drafts in your PR list — the same artifact you'd get from a junior developer who finished a ticket overnight.

The change isn't "agents can now git push." It's that the thing you review moved from a mid-run permission dialog to an end-of-run pull request.

That is the whole insight, and it's worth saying plainly because the mental model is the payoff: you stop supervising the process and start reviewing the result. Code review is a skill founders already have and already trust. Approving forty "is it okay to run git commit" prompts is not a skill; it's a tax.

The safety boundary is structural, not a pinky-promise#

The reason you can let this run unattended is that the guardrail is built into where the work happens, not into the agent's good intentions.

So the failure mode people fear — an agent silently merging half-baked code to production — isn't the mode this ships. The worst case is a draft PR you don't like, which you close. That's a recoverable, familiar outcome, which is exactly why it's safe to scale.

The workflow to actually adopt#

Treat it like staffing, not automation. Three moves:

  1. Scope narrowly before you dispatch. A background agent is only as good as its brief. "Add a rate-limit header to the three write endpoints" produces a reviewable draft; "improve the API" produces a mess you'll close. Write the ticket you'd hand a contractor.
  2. Fan out into worktrees, then step away. Dispatch two or three independent tasks — ones that won't touch the same files — each into its own worktree. Worktree isolation is what lets them run in parallel without stepping on each other; we walked through the mechanics in git worktrees for parallel AI agents.
  3. Review the queue, don't watch the run. When you come back, the drafts are waiting. Review them the way you review any PR — read the diff, run the checks, merge or close. If you still want live visibility mid-run, how to watch background Claude Code agents covers the monitoring surface.

One more detail from the following week's release (v2.1.202–206, July 6–10) that matters for trust: background task notifications now explicitly state that no human input has occurred, so a notification can't be misread as your approval — and sessions that edit, merge, comment on, or push to an existing PR now link it in claude agents. The whole system is being tuned so the review gate stays a human gate.

What it means#

If you run a company of one, your scarcest input is uninterrupted attention. The old background-agent loop spent it in the worst place — a stream of tiny confirmations while the real work sat blocked. The v2.1.198 change moves your one scarce input to the one place it compounds: reviewing finished code. Batch the tasks, let the drafts pile up, and review them like the PR queue they now are.

That's the same lever behind the other July guardrail work across coding agents — the tools stopped shipping horsepower and started shipping where the human sits in the loop. Here, the human sits at the merge button, and only there. For more on that shift, see Claude Code's July 2026 releases: stacked skills and pause-by-default and how the field compares in Devin vs Codex vs Cursor vs Jules: background agents.