Claude Code Now Coordinates Its Own Parallel Sessions. I've Been Doing That Job By Hand.

Our CLAUDE.md has a standing rule that exists because I kept stepping on my own work: "Concurrent Claude sessions share this working tree and git index, so commit with an explicit pathspec." I wrote that after running two or three Claude Code sessions on the same repo at once, each fixing something different, and watching one session's git add -A sweep up another session's staged files. The fix was a rule I had to remember to follow every time. That's the whole problem with running Claude in parallel: someone has to be the coordinator, and until this week that someone was me, manually, with a rule in a markdown file as the only safety net.

On September 17, Anthropic rebuilt Claude Code Projects to actually be that coordinator. Not a folder of reference files like the old claude.ai Projects. A project is now one long-running conversation where Claude decides what becomes a task, starts a cloud session to run it, tracks every session at once, and tells you what needs your attention when you check back. I've spent a few days pointing it at real backlog work on this codebase, and it's worth being precise about what it changes and what it doesn't.

What a project actually is

Strip the announcement language and there are three moving parts. The conversation is where you describe work; Claude acts as coordinator there and doesn't touch code itself. A thread is a full cloud session, Claude Code running on Anthropic's infrastructure with its own branch, its own context window, and its own pull request when the work calls for one. And Overview is the pane that groups every thread by state so you're not clicking into six sessions to find the one that's stuck.

That "one thread per repo" example is Anthropic's own, and it's a real use case, not a demo. Every new thread starts from the same repositories, the same project instructions, and the same project memory, a MEMORY.md file Claude writes to itself as work happens and every new thread reads on start. Correct a thread once and tell Claude to remember it, and the next thread starts already knowing. That's the actual unlock. It's not that threads run in parallel, you could already do that with claude --cloud fired off three times. It's that something is watching all three, holding the shared context between them, and telling you which one needs you back.

The part that costs you

A project isn't a free multiplier on your plan. A new project defaults every thread to Opus at high effort, which is the fastest possible way to burn through a Pro or Max allotment, and there's no soft cap on how many threads Claude starts at once beyond your own instruction to slow down, which the docs are honest is a preference, not an enforced limit. The enforced number is 200 new threads a day across your projects, which sounds enormous until you remember each one is a full session.

Two other limits are worth knowing before you build a workflow around this. A project belongs to one person, full stop, during the beta: no sharing a project or its threads with a teammate, no organization-level visibility. And it's Pro and Max only, in the browser, the desktop app, and the mobile app, rolling out gradually to accounts that have already used cloud sessions. If you don't see Projects in the sidebar at claude.ai/code yet, that's the rollout, not something broken on your end, and there's a waitlist if you want to flag interest.

Where it sits next to everything else Anthropic shipped this year

If you've lost track of which "run Claude in parallel" feature does what, you're not alone. Subagents offload one task inside a single session. Worktrees give parallel local sessions their own working copy so they don't collide on your machine, the exact problem our CLAUDE.md rule exists to paper over. Agent teams spin up teammates for one task and wrap up when it's done. A project is the first of these that persists: it's a standing conversation you keep feeding, not a task that starts and finishes.

For a solo operator or a small team, that's the actual pitch: not "Claude can do more things at once," you already had that, but "Claude will hold the thread between them so you don't have to be the one remembering which session found the bug and which one is still waiting on you." I'm running one against our own drip-email and ad-ops backlog now, repositories plus a standing instruction never to touch the B2B exclusion logic without asking first. Three days in, the thing I actually notice is smaller than "parallel agents": I stopped writing myself Slack messages to remember which session I left mid-task.

Sources: Let Claude coordinate ongoing work with Projects · Use Claude Code in the cloud · Projects redesigned: from folder to conversation (Anthropic)