Anthropic Launches Claude Code Projects in Beta: Parallel Cloud Threads That Keep Working After You Close the Laptop

Anthropic launched Claude Code Projects in beta on September 17, and it is the most explicit move toward multi-agent software development inside a commercial product so far. The old Project was a folder: some files plus one chat. The new one is a single ongoing conversation where Claude acts as a coordinator.

You describe the work; Claude decides what becomes a thread. Each thread is a full Claude Code cloud session running on its own branch with its own copy of the repository. Threads run in parallel, report back to the conversation, and keep going after you close your laptop. Anthropic's launch post describes it as one conversation that splits itself into parallel cloud sessions.

Coordinator and threads

The architecture has two layers. The coordinator reads what you send, answers quick questions in place, and starts threads for actual work — seeing what threads report back, not every step they take. Threads are the workers: a thread opens a pull request when the work calls for one, then watches that PR with auto-fix enabled, pushing fixes when CI fails and replying when checks pass. Threads delegate too, each able to split its assignment further with subagents, loops, and workflows.

Anthropic's own examples are instructive about where this fits: set a goal to reduce checkout p75 latency, then ask Claude to profile each endpoint, test optimizations, and open pull requests in parallel threads. Or retire a deprecated v1 endpoint across API, web, and mobile repositories — one thread per repo, with Claude reporting which PRs merge first. When two threads touch the same code, the overlap surfaces as an ordinary git merge conflict.

What every thread inherits

Standing context is set once and reaches each new thread: the project's repositories and uploaded files, project instructions of up to 16,000 characters, and project memory that Claude writes and reads through a MEMORY.md index — the kind of notes that say the release moved to Friday, or who to check with before touching billing. Each thread also clones every project repository and loads CLAUDE.md, skills, and plugins from all of them. Permission rules, hooks, and env apply only from the directory a thread starts in, so a single-repo project honors them and a multi-repo project does not. MCP tools arrive through the connectors on your claude.ai account, and the project conversation itself has no connectors — connector work has to go to a thread. The Overview pane groups threads by state — Ready for review, Waiting on you, Working, Landing, Idle, Resolved — and a Library tab collects uploaded files and files the threads produced.

The catch: every thread is a session

Anthropic is blunt about the economics: a project uses plan limits faster than a single session, because every running thread is a full session. It exposes a per-project Usage tab plus separate model and effort settings for the coordinator and for threads. By default a new project runs Opus everywhere, with high effort for threads and low effort for the conversation — and users can assign different models or effort levels to the two layers to tune cost against quality. Idle threads also wake and spend again when CI fails or a review comment lands. The enforced ceiling is 200 new threads per day across your projects, and a thread that hits a usage limit waits and resumes on its own.

This is not free extra developers. It is optional concurrency for teams that value wall-clock speed over token economy: sprint weeks before deadlines, large refactors with independent modules, prototype spikes where parallel exploration beats sequential polish. It also puts Anthropic shoulder-to-shoulder with Cursor's Projects launch from last week — one coordinator, thousands of subagents — except Anthropic's threads are persistent cloud sessions on real branches, closer to cloud dev environments than ephemeral agents. The beta is limited to select Pro and Max users on web and desktop, not the CLI. Watch the usage dashboard during beta: parallel output means parallel metering, and surprise limit hits mid-sprint are the obvious failure mode.