Claude Code Projects: One Conversation, Many Parallel Agents

Open five coding-agent sessions and you may discover you have given yourself a new job: keeping five agents informed. One finds a breaking change. Another is still coding against the old API. The release date moves, and now you are copying the same update into three chats.

Claude Code Projects is Anthropic's attempt to take that coordination off your plate. You brief one conversation, and Claude manages the parallel work behind it. The question is how much of your attention that arrangement actually gives back.

You give one conversation an ongoing goal. Claude decides how to split the work, starts separate threads in parallel, passes relevant decisions between them and tracks the results. Each thread is a full Claude Code cloud session with its own context, sandbox and Git branch. The work continues after you close your laptop.

The coordinator preserves project-level memory for future threads. You can still open any thread to inspect its transcript, answer an approval request or steer it directly.

Anthropic launched the redesigned Projects experience on September 17, 2026. It is currently a public beta rolling out gradually to Pro and Max subscribers. Beta access started with people who had used Claude Code cloud sessions and did not already have Projects in Claude chat or Cowork. Team and Enterprise support is planned, but not available yet.

A Claude Code Project accumulates threads, decisions and results across a week

Anthropic's launch-readiness demo: one ongoing project conversation with work split into threads. Original announcement and demo.

Claude Code Projects vs Claude Projects: What Changed

Claude has used the word “Projects” before. The older Claude chat feature groups conversations and a knowledge base into a workspace. It does not create worker threads or coordinate Claude Code sessions.

The redesigned version is an orchestration layer. Its structure has three main parts:

• One project conversation acts as the coordinator.
• Threads perform individual tasks as parallel sessions in the cloud.
• Shared instructions, repositories and project memory give new threads the context they need.

The coordinator sees what threads report back, rather than every tool call they make. Each worker has its own context window and works on its own branch. If two workers edit the same code, those branches can still produce a normal merge conflict. Parallelism removes waiting; it does not remove software integration.

How to Use Claude Code Projects

If the beta has reached your account, open Projects at claude.ai/code or in the Code tab of the desktop app and create a Project. Give it a name, then optionally add a goal, GitHub repositories and reference files. Choose or configure the cloud environment that supplies network access, credentials and tools. Add standing instructions for validation and approvals, then begin with one small representative task before sending a larger batch. Anthropic’s Projects documentation recommends checking the first thread’s result before scaling up.

What changes in practice

Claude Code Projects turns related tasks into one durable conversation.

Imagine that you own a checkout service. You can give the Project an ongoing performance goal, then add a latency regression, a dependency upgrade and release-note work as they arrive. Claude may create one thread to profile endpoints, another to test an optimization and another to inspect the upgrade. If you later say the release moved to Friday, the coordinator can retain that decision in project memory and apply it to later work.

That example is illustrative, not a measured case study. The value depends on whether the work separates cleanly and whether each thread has a reliable way to prove it is done.

Four workflows look especially well suited to the model:

Cross-repository changes

A deprecated API may require coordinated updates to backend, web and mobile repositories. A Project can start one thread per repository, let each run its own tests and open separate pull requests, then report the order in which they should land.

A continuous bug and maintenance queue

Instead of opening a new session for every stack trace, keep one Project for a service. New bugs, review requests and small upgrades enter the same conversation. A lesson from one fix can become memory available to the next thread.

Large migrations

A framework or database migration naturally splits into discovery, implementation and verification across multiple areas. Projects can keep those tasks parallel while preserving decisions such as the target version, rollout order and validation rules.

Research and documentation

A Project does not require a repository. Threads can work from uploaded files or Google Drive folders and return documents to a shared Library. This could fit recurring support-ticket analysis, technical due diligence or release documentation, provided the required sources are available to the cloud environment.

The underlying workflow has evidence

I did not find published Projects-specific outcome measurements. The feature was two days old when I researched this article.

There is, however, evidence that the underlying patterns—long-running Claude Code sessions and parallel delegation—can handle substantial engineering work.

At Rakuten, a Claude Code session implemented an activation-vector extraction method in the 12.5-million-line vLLM codebase over seven hours and reached 99.9% numerical accuracy against the reference method. Rakuten also reported reducing average feature delivery from 24 working days to five, while engineers ran several Claude Code tasks in parallel. Those results cover Claude Code adoption broadly, not the new Projects beta.

Wiz reported using Claude Code to migrate a 50,000-line Python library to Go in roughly 20 hours. The replacement ran at least twice as fast in production, and the company said its top 100 contributors saw a 1.5× increase in merged pull requests. Wiz developers were already running multiple Claude Code instances on separate features. Projects could reduce the manual coordination around that pattern, but I did not find published evidence that it has done so yet.

Claude Code Projects usage limits

Every thread consumes usage like a full Claude Code session, and the coordinator consumes tokens too. Several threads can therefore exhaust a Pro or Max allowance much faster than one session. Anthropic’s help center explicitly warns that running several threads at once uses plan capacity faster. You can reduce the load by choosing smaller models or lower effort levels and asking Claude to run fewer threads.

The beta also has platform constraints:

• Threads run only as cloud sessions with Anthropic as the model provider. Local project threads are planned but are not available today. Cloud sessions continue after you close your laptop and can be steered from another device, as described in the cloud sessions reference.
• Project code repositories currently need to be on GitHub.com with push access and the Claude GitHub App installed.
• Tools, credentials and services that exist only on your laptop are unavailable. A cloud environment controls network access, environment variables, credentials and setup scripts, so private services must be made reachable deliberately.
• A Project belongs to one person. It cannot be shared, and there are no organization-level Project controls during the beta.
• Projects are available on the web, desktop and Claude mobile apps, but not in the terminal CLI.

These are significant limits for regulated teams, private networks and repositories outside GitHub. They also make Projects a poor fit for one small task, a workflow that depends on local hardware, or a job where two agents would constantly edit the same files.

From coding agent to engineering queue

The interesting shift is not simply “more agents.” Developers could already open several sessions.

Claude Code Projects gives that parallel work a persistent coordinator and a shared memory. That makes the unit of interaction larger: instead of prompting one agent to complete one task, you maintain a conversation around a goal and feed it work over time.

For a technical founder, that might be one Project for a launch: backend changes, web polish, release notes and dependency cleanup moving in parallel. For a platform engineer, it might be an upgrade across a family of services. For a developer, it could be the maintenance queue for the system they own.

Claude Code Projects still requires the same discipline as any autonomous coding workflow: narrow tasks, explicit validation, controlled permissions and human review before merge. What it removes is much of the clerical work between sessions. If Anthropic can make that coordination reliable, Claude Code Projects may be less like a new chat feature and more like the first practical inbox for an AI engineering team.