Claude Code Projects turns one engineering goal into parallel cloud-agent branches

Anthropic redesigned Claude Code Projects on September 17, 2026. A project is now a persistent coordinator conversation rather than primarily a folder of files and chats. The coordinator scopes requests, delegates work to parallel threads, reviews results and assembles the outcome. Each thread is a full Claude Code cloud session working against its own branch and repository copy; threads can themselves use subagents, loops and workflows. All threads can contribute to and read shared project memory, while a library stores user files and generated artifacts. The beta began with select Pro and Max users using cloud sessions, with broader plan support promised later.

Coding-agent users increasingly hit a coordination problem rather than a raw model-capability problem: several sessions can work at once, but somebody still has to divide the work, keep context consistent, track branches and combine results. Claude Code Projects moves that orchestration into the product. For builders, this makes multi-agent development accessible without writing a custom supervisor, but it does not remove software-engineering constraints: overlapping edits still create merge conflicts, every thread consumes plan usage as a full session, and the launch version runs in Anthropic's cloud rather than behind the developer's own network.

A developer can describe a larger engineering goal to the project conversation instead of manually opening several unrelated Claude Code sessions. The coordinator decides whether work belongs in an existing thread or needs a new one, delegates tasks and follows the resulting work. That changes Projects from passive context storage into an active orchestration layer.

Each worker thread runs as a full Claude Code cloud session with its own branch and repository copy. Connected threads can run tests and open pull requests independently. Anthropic has not invented a special conflict-resolution model for overlapping edits: if two threads collide, the result is handled as a normal merge conflict. That is a useful boundary because teams can apply existing review and CI practices rather than trusting silent agent reconciliation.

Threads add to and read from shared project memory, allowing decisions such as changed release dates, dropped features or ownership notes to survive across sessions. The design reduces repeated prompting, but it also makes agent-written memory part of the project's operational state. Teams should treat important remembered decisions as inspectable working context rather than an authoritative source of truth.

The product does not make concurrent agents economically free. Anthropic warns that Projects can hit usage limits faster because each thread is a full cloud session. Users can inspect project-specific usage and choose model and effort levels separately for the coordinator and worker threads. That makes orchestration strategy a cost decision as well as a workflow decision.

At launch the worker threads run in Anthropic's cloud. Anthropic says local execution alongside local tools and code and behind the user's network is coming later. Until that arrives, teams with code-location, credential or network-policy constraints need to evaluate whether the beta's remote execution model fits their environment.