On September 17, Anthropic released a redesigned Claude Code Projects in beta, replacing the old multi-session workflow where developers split work by hand, managed handoffs manually and recombined results at the end. In a post titled 'Projects redesigned: from folder to conversation' on the Claude blog, the company said the user now describes the goal and Claude handles the work.
[1][2]The new Projects consists of threads that do the work and a coordinator that directs them. After a user picks a goal and attaches a repository or other context, the coordinator scopes the request, delegates tasks to new or existing threads, reviews outputs and assembles the finished result. The main project chat is the single command point, though users can open an individual thread to inspect and steer details. Anthropic compares the interaction to briefing a chief of staff, who then routes assignments to new or pre-existing threads, checks in and follows through. Users can guide progress from a phone, and work continues after the laptop closes.
Each thread is a full Claude Code cloud session running on its own branch and its own copy of the repository, and can subdivide its assignment through subagents, loops and workflows. Threads draw on and write to shared project memory (files indexed by a MEMORY.md file) plus a library of uploaded files and artifacts, which Anthropic says reduces the need for complex prompt engineering - Claude can recall that a release moved to Friday, why an export was dropped, or who to check with before touching the billing service. When threads edit the same code, the overlap surfaces as a normal merge conflict rather than a silent overwrite, resolved the way any two pull requests collide. Separate branches and repository copies are part of the implementation, not a promise that parallel agents will avoid conflicting edits.
The defaults and limits are explicit. New project threads run on Opus at high effort by default while the coordinator conversation runs at low effort, a configuration Anthropic says draws on a subscriber's plan fastest. Threads currently run only in the cloud; local file and tool access is described as coming very soon, which leaves development tied to local environments and private networks out of the launch. Because each thread is a full session, parallel work consumes usage allowances faster, and the only enforced ceiling is 200 new threads per day per project; a limit a user asks Claude to respect is stored as a preference rather than a hard cap, and idle threads watching a pull request can wake and burn more of the plan when a test fails or a reviewer comments. A Usage tab breaks consumption down by thread and model.
The beta opens to select Pro and Max subscribers who use cloud sessions and have no existing web or desktop projects, widening over the following week, with Team and Enterprise later. Existing projects keep working as before during the rollout.
This is a step from single-session assistance toward a supervision layer for fleets of coding agents, and the open problems are concrete. Merge conflicts between parallel agents still need a human. Cloud-only execution leaves local codebases out, and the shared memory is not inspectable or exportable, a gap for teams that need audit trails. Anthropic's own Frontier Red Team reported in August that agents given conflicting instructions on one project start a turf war, and the redesign hands orchestration to the model while absorbing that coordination risk into the product. Whether the coordinator reliably resolves logical conflicts in code, not just syntactic ones, is the question the beta exists to answer.
[1][2]