Anthropic Rebuilds Claude Code Projects to Run Parallel AI Agent Threads

Anthropic has overhauled the Projects feature in Claude Code, introducing a system where users describe a high-level goal and a coordinator agent splits the work across multiple parallel "threads." Each thread runs as an independent cloud session, capable of opening pull requests and running tests on its own.

What changed

Under the new design, a single coordinator manages task decomposition rather than requiring developers to manually break work into separate prompts or sessions. Progress on each thread is trackable from the main chat interface or individually, including on mobile devices. Over time, Claude builds shared memory across threads within a project, and a library aggregates all uploaded files and generated results in one place.

The feature is currently in beta and limited to select Pro and Max subscribers using cloud sessions in Claude Code, according to Anthropic. Team and Enterprise access is planned for a later date, and local execution — running threads outside Anthropic's cloud — is described as "coming soon." Interested users can join a waitlist, but Anthropic has not disclosed a broader rollout timeline or specific pricing tied to the parallel-thread capability.

Part of a broader push toward autonomy

This update follows Anthropic's recent move to make "autopilot mode" the default behavior in Claude Code, a change the company said improved performance on safety-related coding tasks compared to human developers, according to Anthropic. Neither claim has been independently verified with published benchmark data.

Taken together, the changes reflect a consistent strategy: reduce the number of manual steps a developer takes to get code written, tested, and shipped, while increasing the amount of autonomous computation Claude Code performs per task.

What this means

Parallel agent threads mean more simultaneous model calls per user session, which mechanically increases token consumption — the core unit Anthropic bills against. Framing this as a productivity feature is fair; it likely does cut wall-clock time for multi-part coding tasks. But the shift also moves control over compute spend further from the user and toward Anthropic's infrastructure defaults, since a coordinator — not the developer — decides how many parallel threads a task requires.

For engineering teams evaluating Claude Code, the practical questions are unanswered by this announcement: there's no published data on how threads are billed relative to single-session usage, no benchmark comparing coordinator-split tasks to manually decomposed ones, and no timeline for Team or Enterprise access, which is where much of Anthropic's revenue growth is expected to come from. Until those details surface, this reads as a beta feature aimed at power users on Pro and Max plans, not yet a finished product for organizations managing large codebases.