Anthropic's rebuilt Claude Code projects orchestrate agents only inside cloud sessions

The ZDNET reviewer cannot use the feature he is writing about. He keeps each app he develops in its own directory and has built an orchestration workflow that passes certain messages between those directories, and beyond that the code structures stay isolated ^[14]. All of his Claude Code work runs in a terminal against local files. That rules him out ^[15]. Anthropic's eligibility line names cloud sessions, so a local-only setup does not qualify ^[3][17].

The offer goes to select Pro and Max subscribers who use cloud sessions and "don't have any existing projects on the web and desktop," Anthropic said ^[3]. Subscribers already on projects keep what they have: "If you are a Pro and Max subscriber already using projects, they stay as they are for now, and we'll upgrade them later," the company said ^[4]. The earliest adopters of projects are the ones waiting, behind subscribers who never opened one ^[18].

For anyone through both gates, the pitch is a coordinator. "Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result," Anthropic said. "You can steer progress throughout, and it keeps working after you close your laptop." ^[2] Each discussion becomes a full cloud session with its own branch and copy of the repo, and the coordinator routes new instructions into existing threads ^[5]. Anthropic said to brief Claude on a project "the way you'd brief a chief of staff, and it routes work to new or pre-existing threads" ^[16].

ZDNET has been running swarms of subagents for a while, and Anthropic's announcement said that architecture is what the new projects use internally ^[10]. The change is persistence: the split survives, and you can reopen the discussion belonging to a given subagent and talk to it ^[11]. Threads also seem to start on a button, one Start per thread Claude recommends ^[12]. Naming a thread is what makes mid-flight correction possible, since a run that is visibly going wrong can be redirected before it burns tokens to completion ^[6].

"Every thread now adds to and draws from a shared memory, reducing the need for complex prompt engineering," Anthropic said ^[7]. That memory and the per-thread repo copies live in the cloud today; the company said running "on your machine alongside your local tools and code and behind your network is coming very soon" ^[13]. Work can span repos, with a thread tied to each and feedback on which pull requests to merge in what order ^[8]. It carries on with the laptop shut, and Anthropic said mobile check-ins are on the way ^[9].

It comes down to whether the work already fans out across more than one repo and outlives a single session, and whether a copy of that repo is allowed to sit in a cloud session. Both true, and a pilot on a clean account is cheap. Either false, and there is nothing to migrate to until on-machine threads ship ^[13]. ZDNET's reviewer concludes that the redesign can be game-changing for some classes of work and essentially irrelevant for others ^[19]. In a pilot, count how many threads someone reopened and redirected after hitting Start. If that count sits near zero, the team paid for persistence and used the fan-out it already had ^[11][12].