New Claude Projects: Cloud Threads That Keep Working After You Close Your Laptop

New Claude Projects: Cloud Threads That Keep Working After You Close Your Laptop

Anthropic rebuilt the Projects feature in Claude on September 17, 2026, from a place to store documents and ask questions into an interface for delegating work in bulk. Instead of one conversation working through your task, Claude now splits the work into multiple parallel "threads" that run in the cloud and report their results back to a single conversation. Access, as of September 18, was still limited to a beta, but the rollout is gradual — and the format change is significant enough that it is worth understanding how it works before it reaches your account.

What Has Actually Changed

The old Projects feature was, in practice, a folder: you uploaded files, asked questions about them, and got answers in the conversation. The new version, described in Anthropic's announcement and reported in Japanese trade coverage, flips that logic. Now you describe a work assignment, and Claude breaks it down into parts and launches so-called "threads" — units of work, each handling one task and running in the cloud (freee✕AI consultant / Yosuke, source 6d561f45).

The threads run in parallel, and the results are collected in one conversation. Two practical consequences follow: the work continues after you close your laptop, and you can check progress from the mobile apps. This is no longer a question-and-answer tool, but a queue of delegated jobs that report back.

What Happens Under the Hood

Each thread runs as a cloud session in Claude Code — Anthropic's developer tool. That means each thread gets its own branch and its own copy of the code repository it is working in (AI仕事設計室, source 7509a6e1). With separate copies, the threads can work independently of one another.

But isolation has a price when assignments overlap. If two threads modify the same code, the changes must be merged manually, and the collisions are resolved as ordinary merge conflicts — that is, what developers otherwise handle when two branches in a code project need to be reconciled (source 7509a6e1). For those unfamiliar with version control, this is among the most concrete friction points in the new format.

How Memory Is Shared Between Threads

The threads share a common memory that is stored as files. According to the Claude Code documentation, as cited in the coverage, the memory starts with an index file called MEMORY.md, and its contents can be read, edited, and deleted from the settings screen (source 6d561f45). It is worth noting: because the memory is shared, it is the common context that all threads read — and because it is file-based and visible in the settings, you can go in yourself and correct what the model has incorrectly remembered, instead of re-explaining it in every conversation.

Who Gets Access — and Who Has to Wait

As of September 18, 2026, Anthropic launched the updated Projects as a beta for a limited group:

"Starting today, updated projects are available in beta to select Claude Pro and Max subscribers who are using cloud sessions in Claude Code and who do not have any existing projects on web or desktop."

Quoted from Anthropic's announcement, as reported in the coverage from freee✕AI consultant / Yosuke (source 6d561f45).

Three conditions stand out: you must be a Pro or Max subscriber, you must be using cloud sessions in Claude Code, and you must not have existing projects in the old format on web or desktop. Existing projects continue on the old format, and the rollout is happening gradually — so there is no guarantee you will see the new interface today, even if you meet all the conditions. This coverage is based on secondary reports of the announcement, not Anthropic's primary source, so the details of the access terms may change as the beta is developed further.

Which Tasks Suit Parallel Delegation

Parallel delegation does not pay off for everything. AI仕事設計室 has set out a litmus test for deciding whether a task can be split into threads — and it is worth emphasizing that this is the author's own framework, not an official guideline from Anthropic (source 7509a6e1). The three conditions are:

1. No shared prerequisites. If the threads need the results of each other's work, they cannot run in parallel. The task must be divisible so that each thread can start independently.
2. Completion criteria writable in one sentence. If you cannot formulate when the job is "done" in a single, clear sentence, the thread has no goal to work toward, and you will get results you cannot evaluate.
3. A defined rollback plan on failure. You need to know what you will do if a thread's work turns out to be unusable — otherwise you risk being left with half-finished changes nobody will accept.

The test is a useful translation of merge-conflict logic into practical work division: what makes code branches safe to split also makes work assignments suitable for parallel execution.

The Costs and Limitations

The coverage points to four concrete considerations (AI活用研究班, source d67b9548):

• Faster quota consumption. Multiple threads working at the same time use your usage quota faster than a single conversation. For those working within a subscription cap, this is a direct economic limit on how much you can delegate.
• Merge conflicts. Threads that touch the same code create conflicts that must be resolved manually, as described above.
• Beta status. As of September 18, the feature was only available to a selection of Pro and Max users.
• No access to local machines or corporate networks. The threads run in the cloud and cannot reach your local PC or internal corporate networks directly. Tasks that require files on your machine or access to internal systems must be prepared differently before delegation.

Open Questions to Watch

The rollout is underway, but much remains unsettled. How quickly the beta will expand beyond the first selected users, and how existing projects might eventually be migrated to the new format, are not known from the available sources. It is also not specified how quota consumption is balanced against subscription limits in practice, or whether merge handling will be simplified for users without experience with version control. Most fundamentally: since all of this rests on secondary reports of Anthropic's announcement, it is worth checking the details against Anthropic's own publications as the beta matures. For now, the safe impression is that Projects has changed roles — from folder to work queue — and that those who get early access should start with tasks that pass the litmus test.