Anthropic Rebuilds Claude Code Projects Around Parallel Agents
Anthropic’s redesigned Projects let Claude coordinate cloud coding sessions, shared memory and pull requests across repositories.
What changed
Anthropic has redesigned Claude Code Projects as a coordination layer for long-running software work. Instead of treating a project as a folder of files and a single conversation, the beta lets users describe a goal while Claude breaks the work into parallel threads.
Each thread runs as a separate Claude Code cloud session with its own repository branch and working copy. The coordinator can assign tasks, inspect results, resolve dependencies and assemble the output. A project targeting checkout latency, for example, could ask different threads to profile endpoints, test optimizations and prepare pull requests simultaneously.
The system also introduces shared project memory. Decisions, constraints and working preferences accumulate across threads, allowing Claude to recall details such as a postponed release or a service owner who must approve a change. Projects include a library for uploaded files and generated artifacts, while users can monitor or redirect work from the main conversation, including from a phone.
Access and limits
The feature launched in beta for selected Pro and Max subscribers using cloud sessions in Claude Code. Existing Projects remain available during the rollout, with wider access planned for Claude, Team and Enterprise users. Threads currently operate in Anthropic’s cloud and cannot reach files or tools on a user’s local network.
Parallel execution can also consume usage limits faster. Shared memory and branch-based work reduce coordination overhead, but merge conflicts remain ordinary engineering problems when agents touch the same code.
Why it matters
The launch moves coding agents toward project-level delegation rather than turn-by-turn assistance. Its significance is less about adding another coding interface than about making multiple autonomous workstreams a normal unit of software production. The unresolved question is whether coordination quality, review burden and usage economics improve enough to justify giving an orchestrator control over several agents at once.