Anthropic puts a coordinator above Claude Code agents

Until now, continuing a project across several Claude Code sessions required manual bookkeeping. Developers had to maintain a separate status record and provide it to Claude again, or ask Claude Code to reconstruct the previous conversation and determine what had already been done.

Claude Code Projects is designed to take over that reconstruction. A project stores a shared goal and context, creates separate work streams for different parts of the assignment, and keeps a shared memory that is updated as the work progresses. Anthropic describes the architecture simply: the streams do the work, while Claude manages them.

That model fits systems that are never really finished: digital storefronts, websites, content management systems, procurement platforms and other business applications. Their engineering teams spend years changing, fixing and extending them rather than delivering a final version and moving on.

A developer might create a project to reduce the p75 checkout latency of an application. Claude could create parallel streams to profile endpoints, test optimizations and open pull requests. In another example, a team connects the repositories for its API, web and mobile applications and asks Claude to remove a deprecated version 1 endpoint. Separate streams migrate the code that calls it and test each repository, while the coordinator tracks dependencies and tells the developer which changes should be merged first.

Each stream is a separate cloud Claude Code session with its own branch and repository copy. Streams can also delegate work to subagents, loops and workflows. The arrangement does not remove ordinary engineering problems: when streams edit overlapping code, standard merge conflicts can still occur and must be resolved in the usual way.

The developer does not need to wait for one stream to finish before starting another. New instructions can be sent to an existing stream or used to create an additional one. A developer can open an individual stream to inspect or redirect its work, while an overview panel shows tasks that need human attention.

Anthropic says developers will also be able to manage individual streams from a phone. The sessions continue running in the cloud after the developer closes a laptop or loses a network connection, although the current cloud environment cannot access files, tools or systems on the developer’s computer or inside an organization’s internal network.

The feature is easy to confuse with Claude Projects, which Anthropic first introduced in June 2024. That product is a general-purpose Claude workspace for grouping chats, files, instructions and project knowledge. Users can hold multiple conversations and preserve shared knowledge, but each chat remains a separate conversation unless information is explicitly added to the project’s knowledge or memory.

Claude Code Projects, introduced on September 17, addresses a different problem. It coordinates AI coding agents across multiple Claude Code sessions and gives them a shared working memory. Decisions made during one part of an initiative — such as moving a release date, rejecting a feature or consulting a particular person before changing a service — can influence later work elsewhere in the project.

The distinction is more than branding. Claude Projects helps Claude remember context supplied by the user. Claude Code Projects is meant to help Claude remember and coordinate work that Claude itself has carried out.

The beta initially goes to selected Claude Pro and Max subscribers who use cloud Claude Code sessions and do not yet have Projects in the web or desktop version. Anthropic says it will expand access to more Claude Code users on those plans during the following week, then bring the updated format to a broader Claude audience, including Team and Enterprise. Existing Claude Projects on Pro and Max will continue to work as before during the rollout.

This puts Anthropic into a contest that is moving beyond the quality of an individual coding agent. OpenAI, GitHub and Google are also building asynchronous, parallel and longer-running agent workflows. The common capabilities are increasingly clear: isolated environments, multiple agents, persistent context and some form of human oversight.

The more consequential difference is where each company places control. OpenAI’s Codex is framed as a collaborative coding workspace where several agents can run in parallel, isolate their changes and support longer-running work. GitHub anchors its agent tools in the repository, linking sessions to branches, issues, pull requests and continuous-integration workflows. Google’s Jules supports asynchronous development and stores memory at the repository level.

Anthropic is placing a conversational coordination layer above individual sessions and, in its example, above several repositories. The developer continues talking to one Claude Code project, while Claude decides whether a new instruction belongs in an existing stream or requires a new one. Its shared memory is intended to cover the broader initiative rather than just a repository or a single conversation.

That is a narrower and more credible claim than saying Anthropic has invented parallel coding. It has not. The notable bet is that the main interface for multi-agent development should be one continuing conversation with a coordinator, rather than a repository dashboard or a collection of agent windows.

The announcement is quieter about how that coordinator’s memory should be governed as projects grow. Anthropic explains that decisions and materials can persist, and that developers can tune the frequency of messages, new streams and update detail. But the central operational problem remains: someone must know which remembered decision is still valid, which requirement has changed and which agent’s work can safely be merged. The product moves that burden into Claude; it does not make the burden disappear.

There are also practical constraints. Claude Code Projects does not charge separately. Projects consume the subscriber’s existing Claude allowance, and every stream is a full Claude Code session. Running several streams at once can therefore exhaust usage limits faster. Users will be able to see consumption by project and choose models and effort levels for the coordinator and work streams.

Local support is the next important boundary. Anthropic says it will add it “very soon,” allowing Claude Code Projects to work with local code and tools, including systems inside an organization’s network. Until then, cloud execution is useful for persistence and remote supervision, but it excludes the environments that many companies cannot expose to a cloud session.

Claude Code Projects therefore makes a specific trade: it treats the project conversation as the control plane and the coding sessions as replaceable workers. If Anthropic can extend that model into local environments without losing the shared memory and simple supervision it is promising, the coordinator may become the part developers notice. For now, the feature’s strongest idea is also its unresolved dependency: long-running agents need an enduring manager, but that manager must remain understandable to the human responsible for the result.