Anthropic has fundamentally overhauled the projects feature in Claude Code. Instead of working through one task after another, a coordinator now takes over the distribution: it breaks down a requirement and passes the subtasks to several cloud threads that run in parallel. These threads work independently, open pull requests – i.e. proposed code changes that are reviewed before being merged – and run tests.

So that the parallel strands do not work past each other, according to Anthropic they draw on a shared memory. Insights from one thread are thus also available to the others. The feature is in beta and is initially available to selected Pro and Max subscribers.

What changes in the working model

Until now, working with AI assistants in development was predominantly dialogic: a person describes a task, checks the result, corrects it. The new model shifts this process. The description of the task remains the central point of intervention, but execution is split up and continues concurrently without every step being confirmed individually.

In practice this means: the bottleneck moves. It is no longer the writing of code that sets the pace, but the quality of the task description at the beginning and the review of the results at the end. Anyone receiving five pull requests at once needs a review process that can handle this – otherwise a new queue simply forms.

New requirements for quality assurance

The fact that the threads run tests themselves is helpful, but does not replace a robust test strategy. Automated tests check exactly what someone previously defined as worth checking. In projects with thin test coverage, a passed test run therefore provides little assurance. For teams that want to rely more heavily on automated implementation, it is worth investing first in tests and in a reliable build and deployment chain.

At the same time, the need for traceability increases. When several strands work in parallel on one repository, it should remain clear for every change which task it answers and who approved it. Pull requests are a suitable format for this because they bundle the change, the rationale and the review in one place.

What project managers should watch out for

  • Task scoping: The more clearly and distinctly a requirement is described, the less rework arises during parallel execution.
  • Review capacity: More pull requests per day require fixed responsibilities and realistic time windows for review.
  • Test coverage: Automated tests are the prerequisite for fast cycles not coming at the expense of stability.
  • Conflict management: Parallel changes to the same files create merge conflicts. Clear module boundaries in the code reduce this risk.
  • Data protection and confidentiality: Cloud-based execution means that project code is processed outside your own infrastructure. This must be clarified in advance.

Assessment

The announcement concerns a beta with limited access, not a generally available feature. Reliable statements on result quality with larger, grown codebases are not yet available. The direction is nevertheless clear: providers are shifting not only the writing of code but also the coordination of work into the tools.

For web and software projects, this above all means that the focus of the work shifts: away from pure implementation, towards precise specification, clean architecture and consistent review. Those who have already organised these foundations in their project benefit most quickly from shorter implementation cycles. Where tests are missing and responsibilities are unclear, automation above all accelerates the creation of rework.

Sources