Anthropic announced a redesigned Claude Code Projects beta on September 17. A coordinator delegates and reviews work across parallel cloud threads, letting users guide a project through one main conversation.
Who gets access first
The initial beta targets selected Pro and Max subscribers using Claude Code cloud sessions who have no existing projects on web or desktop. Anthropic plans to expand access; existing projects keep working. Team, Enterprise, chat and Cowork support follow later.
Parallel work still needs reconciliation
Each thread has its own repository copy and branch. Anthropic says overlapping changes are resolved as merge conflicts. Shared memory and a library carry context and files between tasks. Users can steer the coordinator or individual threads.
Each thread is a full session, so parallel work can reach usage limits faster. Users can inspect project usage and select models and effort for the coordinator and workers. Threads run in the cloud at launch; local execution is planned.
These are announced capabilities, not independently tested findings. Anthropic provides no benchmark here establishing a productivity gain.
NextWith.ai perspective: measure the handoffs
Our view is that the interesting test is how much coordination work remains for the person supervising the project. Producing several changes at once does not, by itself, establish that the combined result is ready to ship.
A useful trial would start with a small, reversible assignment whose success criteria are written down in advance. Separate tasks should have clear boundaries and an agreed order for integration. Record the time spent inspecting outputs and resolving disagreements, as well as the time spent waiting for generation. This is our proposed evaluation, not a workflow we have tested with this beta.
Watch for a particularly revealing failure: two individually plausible outputs that rely on incompatible assumptions. That is where a coordinator must do more than summarize progress. The practical question is whether it identifies the disagreement early enough to reduce human rework.
For a team considering adoption, a finished, verified change is the meaningful unit of output. Counting agents or completed threads alone would miss the effort required to turn parallel activity into a usable result.
