Anthropic redesigns Claude Code Projects: from a file folder to a conversation that coordinates the work

Anthropic has unveiled a redesign of "projects" in Claude Code that changes their very nature: they are no longer a passive folder where files and scattered conversations pile up, but a workspace where the user describes a goal and Claude takes care of organizing, delegating and assembling all the work needed to achieve it. At the heart of the change are so-called "threads", which carry out specific tasks, and a "coordinator" that directs them, reviews what they produce and presents the final result.

It works as follows: when starting a project, you choose a goal along with the repository or working context, and Claude immediately proposes tasks it can begin to tackle. The user can configure the project's cloud environment, the connectors, the plugins, the instructions and the model to be used. From there, progress can be followed from the project's main chat, or you can go into each individual thread to review and fine-tune the details. The article compares this way of giving instructions to briefing a chief of staff: you set the general direction and Claude distributes the work across new or existing threads, tracking it and verifying that it gets completed.

As concrete examples, the article mentions two scenarios: setting up a project with the goal of reducing an application's p75 checkout latency, asking Claude to profile each endpoint, test optimizations and open pull requests in parallel threads; and connecting the API, web and mobile repositories with the goal of retiring a deprecated v1 endpoint, so that Claude creates one thread per repository to migrate the calls, run the tests, open the corresponding PRs and finally indicate which ones should be merged first.

On the technical side, each thread is in fact a full Claude Code session in the cloud, working on its own branch and its own copy of the repository. If several threads touch the same code, the overlap is resolved as a normal merge conflict, just as it would be with any other pull request. In addition, each thread can itself subdivide the delegated work using subagents, loops and workflows when the assigned task is large enough.

Another central element of the redesign is shared memory: over time, Claude learns details about the project and applies them in later work, so that every new thread draws on that common memory and contributes to it. The article gives concrete examples of the kind of information it can remember: that the launch moved to Friday, why an export feature was removed, or who needs to be notified before touching the billing service. As explained, this reduces the need for complex prompt engineering with every new request. Claude also learns the user's working and communication style, making it possible to adjust how often it reports progress, how frequently it opens new threads, or the level of detail in its updates. Alongside this memory, projects now include a "library" that brings together both the files added by the user and the artifacts generated by Claude, making it easier to find relevant material and build new work on top of what has already been done.

On availability, the article (dated 17 September 2026) notes that this version of projects has been in beta since that day for a selection of Pro and Max subscribers who use cloud sessions within Claude Code and who do not already have existing projects in the web or desktop versions. Over the following week, access would be extended to more users on those same plans, and it would then reach all Claude plans, as well as Team and Enterprise. Those with existing projects on Pro or Max will keep using them exactly as they work today, and they will be updated later as the rollout to chat and to Cowork progresses. For Pro or Max users who do not yet have access, there is the option to join a waiting list.

The article itself points out one relevant practical consequence: because several threads can run at once, and each one is a full Claude Code session, projects can burn through usage limits faster than before. To offset this, users can check each project's specific usage and select both the model and the effort level used, for the coordinating chat as well as for the individual working threads.

Looking ahead, the article reveals that, although threads currently run in the cloud, it will soon be possible to run them on the user's own machine, alongside their local tools and code and within their own network, though no further details or specific date are given for that capability.

Taken together, this change fits into a broader trend towards so-called "agentic coding", in which the user stops micromanaging individual tasks and instead sets high-level goals that an AI system breaks down, runs in parallel and supervises on its own. The design built around threads, a coordinator and persistent memory seeks to solve a real problem in long-running workflows: the loss of context between sessions and the burden of manually coordinating multiple simultaneous workstreams. At the same time, the article openly acknowledges the trade-offs of this approach, in particular the faster consumption of usage limits given that these are full sessions running in parallel, which makes it necessary to provide granular controls over model and effort so the user can balance cost and speed.