Anthropic rolled out a redesigned version of Projects on September 17, starting in Claude Code, where a single conversation now lets Claude break a goal into threads that run in parallel in the cloud. Claude Code Projects is in beta for select Pro and Max subscribers, and the work keeps going after you close your laptop.
Key Takeaways
- A coordinator hands out work to threads, and each thread is its own Claude Code cloud session on its own branch.
- Threads share memory and a common library of files and artifacts, and can open pull requests and run tests.
- Anthropic warns that running several sessions at once burns through usage limits faster.
Have an AI Sum Up This Article
ChatGPTHow Claude Code Projects turns one brief into parallel threads
Anthropic sums up the idea in one line: brief Claude in the project the way you would brief a chief of staff, and it routes the work to new or existing threads. The announcement on the Claude blog lays out a two-layer design, with a coordinator that assigns work and threads that carry it out.
The developer account put it more mechanically. In its launch post on X, the ClaudeDevs team described a project as one conversation that splits the work into threads by itself, runs them as parallel cloud sessions, passes context between them and keeps going when you leave.
The main Claude account framed it for a wider audience in a separate post on X: you describe what needs doing, and Claude directs parallel threads that keep working after you close your laptop. The same post says Claude Code Projects starts in Claude Code and is coming to all Claude users soon.
Every thread in Claude Code Projects is an independent cloud session working on its own branch and its own copy of the repository. When two threads touch the same code, the overlap shows up as an ordinary git merge conflict, which keeps the whole thing inside tooling developers already know.
With a connected repository, threads open pull requests and run tests. They can also delegate further to subagents, loops and workflows, so a single conversation can end up sitting on top of several layers of delegated work.
Anthropic gave two examples of the kind of job it has in mind. The first is cutting checkout latency by profiling endpoints and testing several optimizations in parallel. The second is retiring deprecated API versions across multiple repositories at the same time.
Both are bounded, measurable engineering chores. They lend themselves to parallel work because each thread can own a separate endpoint or a separate repository without stepping on the others, which is exactly where Claude Code Projects looks strongest on paper.
The release follows the autonomy push Anthropic started over the summer, when Claude Code switched to auto mode by default on August 14. This time the agent is coordinating other agents instead of working alone. Claude Code has been adding capabilities all year, including a built-in web browser that lets it act on live sites, and each thread now inherits that growing toolkit.
Shared memory, a common library and usage limits
Threads draw on a shared memory, which Anthropic says cuts down on prompt writing for every new task. Claude also keeps project-specific context over time, such as a schedule change, a design decision or a team protocol. In practice, a thread started on Friday should not need to be told again on Monday that the release date moved or that the team agreed to freeze a given API.
The system adapts to how each user likes to work, including how often it checks in and how much detail it puts in updates. For a lead developer running several workstreams, that matters as much as the parallelism itself, because a project that pings too often defeats the point of delegating.
A built-in library collects the files and artifacts Claude produces along the way. It builds on the artifact sharing Anthropic opened in Claude Code in June, and gives every thread access to the same set of outputs.
Anthropic flags that Claude Code Projects can hit usage limits faster, since several sessions run at once. Users can monitor their consumption and pick which model runs the coordinator and which model runs the worker threads.
That model choice is the main cost lever. A heavier model on the coordinator and a lighter one on the threads is the obvious split to try, although Anthropic has not published guidance on the best combination.
The timing is awkward for subscribers. It lands three days after Claude Code’s weekly limits dropped 17% on September 14, so parallel threads now draw on a smaller allowance than they would have a week ago.
Access is narrow for now. The beta covers select Pro and Max subscribers who use cloud sessions in Claude Code and have no existing projects on the web or desktop apps. Anthropic plans to widen access over the coming week, with Team and Enterprise to follow later.
Execution is cloud-only at launch, and running Claude Code Projects on a local machine is described as coming very soon. Anthropic has also opened a waitlist form for users who are not part of the first wave.
A cloud session runs on Anthropic’s infrastructure instead of the developer’s laptop, which is what lets the threads keep working with the lid closed. The flip side is that the repository has to be reachable from the cloud, a condition some companies with locked-down codebases will need to clear before they can try it.
More articles on Horizon
- OpenAI Models Tried to Hide Their Mistakes
- DeepMind Institute Tackles the Rise of Human-Level AI
- Review a Contract With Claude Without Leaking Data
What Claude Code Projects changes for dev teams and rivals
For a developer, the shift is in what gets handed over. The unit of work becomes a goal, which Claude turns into a plan, a set of branches and a stack of pull requests to review.
Review becomes the bottleneck. Several threads produce several branches, and a human still has to approve merges they did not write, which moves the effort from writing code to reading it. Five parallel branches still have to be read one at a time by the same reviewer.
Teams adopting Claude Code Projects will likely want tighter review rules, such as smaller pull requests per thread or a required test pass before anything reaches a human. Otherwise the speed gained upstream simply piles up as a review queue downstream.
Budget is the other constraint, and engineering leaders already know it well since the bill for AI coding agents started worrying IT departments. Parallel threads deliver faster, but they also multiply consumption on every task.
On the competitive side, Anthropic is putting agent orchestration straight into its flagship developer tool instead of leaving it to third-party frameworks. Rival coding assistants now have to show they can coordinate several sessions with shared memory, in a product developers already use every day.
The move also ties into the rest of Anthropic’s week. The day before, chat and Cowork merged into a single Claude app, and Claude Code Projects applies the same logic to engineering: describe the outcome, and let Claude decide how to split the work.
Early testers would do well to start with bounded jobs like the two examples Anthropic published, and to watch their usage dashboard closely during the first runs. A vague goal spread across many threads is the fastest way to reach a weekly limit with little to show for it.
The clearest winners are teams sitting on a backlog of small, independent maintenance jobs, the kind that rarely gets prioritized because nobody wants to spend a sprint on it. Claude Code Projects can split that backlog across threads and hand back a set of reviewable changes. Solo developers on plans with tight allowances gain less, since every extra thread eats into the same weekly budget.
What comes next depends on two releases: local execution, and the opening to Team and Enterprise plans. That second step is where the cost of Claude Code Projects will be measured across entire teams instead of individual subscriptions.
Follow the story on Horizon.
