Anthropic relaunches Claude Code Projects with parallel agents
Anthropic relaunched Claude Code Projects on September 17, 2026, as a control layer for orchestrating multiple coding agents within Claude Code, according to VentureBeat.
Parallel threads and shared memory
The redesigned product introduces parallel coding threads — each running as a separate Claude Code session on its own git branch — coordinated by a central project agent that maintains shared memory across threads. The Register reported that the architecture allows developers to delegate sub-tasks to independent agent threads while the coordinator tracks progress, dependencies, and project state across all branches.
The shared memory layer stores project context, file artifacts, and long-running task definitions in what Anthropic described as a "cloud-session workflow" — meaning project state persists across user sessions and agent restarts, unlike stateless single-turn interactions.
Access and rollout timeline
The feature is currently available to select Claude Pro and Max subscribers with cloud-session access enabled, according to Ground News. Anthropic has signaled a broader rollout to additional users and subscription tiers in coming weeks, though a firm general-availability date has not been announced.
Local development workflows and offline usage are not yet supported; the coordinator currently requires cloud infrastructure to maintain shared state across parallel threads.
Why it matters for the agent economy
Claude Code Projects targets a concrete gap in agent workflows: task delegation and memory persistence. Prior to this redesign, Claude Code operated in single-thread, stateless sessions — each conversation reset. Developers working on multi-faceted projects (refactoring, testing, debugging in parallel) had to manually re-brief the model on context between conversations.
The parallel-thread model lets developers spin up agents to tackle independent sub-tasks — one thread might write a test suite while another refactors a module — with a supervisor agent monitoring all threads and arbitrating conflicts. This mirrors real engineering team workflows and addresses a key limit in prior agentic coding: sustained, multi-threaded project ownership.
AI Chat Daily and Venture Beat have both identified this as a step toward agent orchestration platforms — software that treats AI models as team members rather than stateless query engines.
The product is concrete and shipping to a bounded user group; it represents an actual capability addition (parallel execution + persistent memory) rather than a roadmap claim.