Claude Code cloud sessions GA: agents run without you

An agent that stops when you close your laptop has a natural safety valve. Remove it and you need a different one. On September 23, Anthropic took Claude Code cloud sessions out of research preview and made them generally available, with a one-time launch credit — $100 on Pro, $250 on Max — for existing subscribers. Claude Code cloud sessions run on Anthropic's infrastructure, which means the work continues after you walk away from the machine.

What actually happened

Anthropic's Claude Code account announced the move out of research preview, describing cloud sessions as keeping Claude Code working "even when your laptop is closed." The credit is claimable in the CLI with /claim-credit, and reporting puts the claim deadline at October 7. Anthropic clarified that the credit is optional and spends first — cloud sessions otherwise draw on your normal Pro or Max plan usage rather than a separate meter.

That is the whole announcement. It is small, and it changes a workflow assumption most teams never wrote down.

Why long-running agents matter for your business

Until now, the laptop was a rate limiter. Your agent ran while you watched it, and when you shut the lid, it stopped. A lot of teams were relying on that without knowing it.

Your review step has to become explicit. When a session can run for hours unattended, "I'll notice if it does something weird" stops being a control. The control is the pull request. Cloud sessions are wired to GitHub, so the natural pattern is: agent works on a branch, opens a PR, a human merges. That is a real gate. Make it the only path to your main branch and the unattended part stops being scary.

Scope the credentials before you scope the task. An agent running on someone else's infrastructure needs its own repo token with its own permissions, not a copy of yours. Write access to one repo, no org-wide scope, no production secrets in the environment. We set this up per project, and it takes about twenty minutes.

The launch credit is a test budget, so use it as one. $100 on Pro buys you enough runtime to find out whether unattended agents actually help your workflow or just generate more PRs to read. Pick the boring jobs first — dependency bumps, test coverage on a neglected module, a migration you have been avoiding. Those have clear pass or fail signals. Anything requiring product judgment is the wrong first test.

The honest read: this is a convenience feature with a governance tail. The convenience is real. The tail is that unattended work needs a gate you actually enforce, and most teams discover which gates are real only after something merges that should not have.

Sources: Claude Code (Anthropic), Claude Code docs.