Claude Code Projects: Parallel AI Coding Agents That Keep Working While You Sleep

Anthropic launches Projects in Claude Code beta: a coordinator directing parallel coding threads with shared memory that keep running in the cloud after you close your laptop.

Claude Code Projects: Parallel AI Coding Agents That Keep Working While You Sleep
Table of contents

Claude Code Projects: Parallel AI Coding Agents That Keep Working While You Sleep

Last updated: September 2026

On September 17, 2026, Anthropic launched a redesigned version of Projects inside Claude Code that changes how developers work with agents: instead of juggling multiple sessions by hand, you describe the goal in natural language and a Claude "coordinator" directs parallel worker threads — each one a full Claude Code cloud session on its own branch and repo copy — that keep running after you close your laptop. The feature is in beta for selected Pro and Max subscribers right now.

Official launch card for Projects in Claude Code

Source: Official Claude blog

Official announcement post about Projects in Claude Code

Source: Claude's official X account

You have lived this problem

Every developer working with AI agents knows the drill: one session for the feature, another for tests, a third for the flaky bug — and most of your day goes into splitting work, tracking handoffs, and stitching results back together. Anthropic compresses the fix into one sentence: "you describe what needs to get done and Claude manages the work." The coordinator scopes the request, delegates the work, runs parallel threads, reviews outputs, and assembles the finished result.

The company's own usage advice captures the philosophy: "Brief Claude in the project the way you'd brief a chief of staff." Your role shifts from task juggler to goal-setter.

How the coordinator and threads actually work

Official diagram of the start-new-threads interface in Projects

Source: Official blog post

The architecture is simple at its core: threads do the work, a coordinator directs them. When you create a project you pick a goal plus repo and context, and Claude suggests where to start. Each thread is a full Claude Code cloud session on its own branch with its own repo copy; if two threads touch the same code, conflicts resolve as a normal pull-request merge — exactly like a human team. Threads can subdivide work further through subagents, loops, and workflows.

Shared memory is the deepest addition: every thread adds to and draws from a shared project memory, alongside a library of files and artifacts. The project remembers its decisions and your communication style — check-in frequency, update detail. You can steer progress from your phone, execution is fully cloud-based for now, and Anthropic promises local execution "alongside your local tools and code and behind your network" is coming very soon.

Two official examples show the scale: cutting checkout p75 latency by profiling endpoints, testing, and opening parallel PRs; and retiring a deprecated v1 endpoint across three repos (API, web, mobile) with one thread per repo — days of coordination compressed into a single request.

Quick comparison: manual session management vs Projects

Aspect Hand-managed sessions New Projects
Work decomposition You analyze and assign everything Coordinator proposes and assigns automatically
Parallel execution Separate sessions needing individual attention Parallel threads on independent branches
Memory Each session's context is isolated Shared, accumulating project memory
Tracking Personal vigilance across windows Overview panel flags what needs attention
Continuity Stops when you close the machine Cloud sessions keep running

What does this mean for you? Getting into the beta

The feature works from anywhere through the cloud — Nigeria, India, the Philippines, Brazil, or anywhere with a stable connection — and requires a Pro or Max subscription. Entry point is claude.com/projects; if you do not see it yet, rollout is gradual over the coming weeks and there is an official Projects waitlist form. Developers who have already used Claude's coding tools will find the learning curve gentle: the idea simply turns a prompt writer into a team manager.

Compared with familiar alternatives: traditional coding assistants complete the line you are typing; single-session agent platforms run one task at a time; Projects manages a portfolio of interrelated tasks with shared memory — a different functional generation, with one important caveat: every thread is a full session drawing from your allowance, and wide parallel runs can burn through usage limits faster than you expect.

Who should try it now — and who should wait

Try it now if you manage decomposable engineering work on an active repository: retiring legacy interfaces, tuning performance on specific paths, migrating libraries, or clearing a backlog of related fixes — exactly the cases Anthropic used in its two official examples, and exactly where a coordinated thread portfolio beats a single session.

Waiting is smarter in three cases: your repositories are sensitive and never leave your network in the first place (wait for the promised local execution), your plan's usage allowance is tight and wide parallel runs would drain it before you see value, or your team needs centralized permission management on Team or Enterprise plans that the first wave has not reached yet.

One professional note for developers everywhere: this is the first time the "agent manager" idea has moved from single tasks to full projects with accumulating memory. Learning to write clear, decomposable goals and to review parallel merges efficiently is building precisely the skill engineering teams will screen for over the next two years — and early hands-on experience here beats any certificate.

Limitations to know before you try it

First, this is a gradual beta: not all Pro and Max users have it yet, and Team and Enterprise plans come later. Second, the faster usage-limit burn is real, not marketing: you can monitor per-project usage and set model and effort levels separately for the coordinator and the worker threads. Third, execution is cloud-only today — anyone working with sensitive repositories behind internal networks waits for the promised local mode. Fourth, the interface and documentation are English-first; prompts in other languages work, but precision is highest with careful English instructions.

What is the difference between new Projects and the old Claude Projects feature?

The older Projects (from 2024) are workspaces for organizing chats, files, and knowledge. The redesigned version is a coordination system for parallel coding threads with a director agent. Existing projects keep working and will be upgraded progressively.

Does Projects run on my local machine?

Not yet. Current execution is cloud-only; Anthropic says local execution "alongside your local tools and code and behind your network" is coming very soon.

Which subscriptions get access today?

The beta is open to a selected group of Pro and Max subscribers using cloud sessions in Claude Code who have no existing projects on web or desktop, expanding over the following week.

How do I avoid burning through usage limits?

Cap the number of parallel threads, pick lower model and effort levels for routine tasks in the coordinator and thread settings, and check the per-project usage view before widening a request.

Can a project wreck my repository?

Each thread works on its own branch and repo copy, and changes merge as normal pull requests that you review — closer to human team practices than direct main-branch edits.

Your practical playbook if you have access today: start with one small project and a single clear goal, let the coordinator propose the breakdown before scaling, tune effort levels, and review the merged result yourself. If your foundations in programming and AI need strengthening first, the free and certified course collections at Truescho are a practical starting point, and our earlier coverage of the Claude Fable 5.1 release and Anthropic merging Cowork into Claude completes the picture.

Sources