Anthropic Adds Parallel AI Worker Coordination to Claude Code Projects, Caps 200 Threads Daily

Anthropic updated Claude Code Projects with coordinator messages, read-only thread sharing, reminders, branching commands, and improved artifact navigation. Each project is a long-running conversation with a coordinator that assigns work to threads. Every thread runs as an independent Claude Code cloud session with its own repository copy and Git branch. The coordinator now posts pending decisions, completed tasks, and blocked work in project chat, reducing manual monitoring of thread lists across parallel cloud sessions.

The update is live on web, desktop, and latest mobile app. Claude Code Projects remain in public beta, introduced September 17, for Pro and Max subscribers on a waitlist, while Team and Enterprise plans are unsupported. Anthropic caps accounts at 200 new threads per day across all projects. Projects carry no separate fee, but parallel workers draw from the same Claude Code allowance, and a thread that hits a limit pauses until capacity returns. New projects default to Opus for coordinator and workers, with low coordinator effort and high worker effort.

The design turns cloud-only projects into a queue of parallel engineering tasks that can be reconciled through Git branches and pull requests. Anthropic lists performance profiling, cross-repository migrations, dependency upgrades, test repairs, and bug triage as fits. Tasks with strict dependencies still require humans to sequence pull requests and resolve conflicts. That boundary matters: parallelism helps when work is decomposable, while small one-branch changes may still be cheaper in a single Claude Code session.

For developers, the change shifts attention from watching sessions to reviewing reports, artifacts, and merge conflicts. For teams, the economics are explicit: more workers consume the same subscription allowance, and high-effort Opus threads can exhaust limits faster. The read-only link rule also shows a security boundary, since threads that touched a local computer cannot be shared, limiting exposure of local files and shell output in cloud transcripts.