Anthropic Rebuilds Claude Code Projects for Parallel Agents
Anthropic has redesigned Projects in Claude Code to run multiple agents against shared memory, shared goals, and a common library of files and artifacts. The updated feature is in beta for select Claude Pro and Max subscribers, The Verge reported on Sept. 17.
Each project runs multiple threads on different tasks in parallel. A coordinator directs the work. Users can work inside any thread individually or stay at the project level, monitoring progress and issuing updates through the main project chat.
Each thread is a Claude Code cloud session. It gets its own branch and its own copy of the repo. That isolation is structural. It lets parallel lines of work proceed without immediate interference, with reconciliation deferred until integration.
When threads edit the same code, the overlap is resolved as a merge conflict, handled like a pull request. Threads can also break their own assignments down further, delegating to subagents and running loops and workflows under the coordinator.
At launch, threads run in the cloud. Support for local tools and local code is coming very soon. Anthropic plans to expand Projects access to all Pro, Max, Team, and Enterprise users, as well as to Cowork and regular Claude chats.
Looking at what this means for working engineers, the shift is from managing one agent to managing concurrency between many. Single-agent assistants left task decomposition, sequencing, and integration to the developer. Here the system takes on more of that scheduling work, while the developer moves toward review and direction.
In my view, the use of branches and PR-style conflicts as the coordination mechanism is a practical choice for that audience. Engineers already reason about isolated branches, diffs, and merges every day. Reusing those primitives for agent output means no new abstraction to learn, and a direct fit with existing review, testing, and CI practices.
The operational details will matter more than the architecture diagram. Conflict frequency under parallel edits, the fidelity of shared memory across threads, and how well coordinator context holds up as thread count grows will determine whether a project feels like a coordinated team or a set of competing drafts. Interaction design will matter too, since the ability to intervene per thread or from the main chat puts the developer in the role of dispatcher as well as reviewer.
If that coordination cost stays low, the payoff is straightforward. Small teams could keep several workstreams moving at once without constant context switching, letting agents explore implementations in parallel while humans focus on intent, interfaces, and acceptance. That is a quiet but useful change in how coding work gets divided.
