# Claude Projects Threads: How Anthropic's New Agentic Coordinator Works > Source: > Published: 2026-09-19 00:00:00+00:00 # Claude Projects Threads: How Anthropic's New Agentic Coordinator Works Anthropic rebuilt Claude Projects into a coordinator that delegates work to cloud threads and sub-agents, with shared memory and RAG retrieval. ## What is the new version of Claude Projects? Claude Projects used to be a folder: instructions, uploaded files, and a memory that every chat inside it could read. The new version, currently in beta for Claude Code, turns a project into a single long-running conversation that acts as a coordinator. Instead of doing the work itself, that top-level conversation breaks a goal into separate threads, hands each one off to a full Claude Code cloud session, and reviews what comes back. Each thread runs on its own branch and repository copy, and can spin out its own sub-agents and workflows underneath it. Anthropic’s framing is blunt: threads do the work, Claude directs it. ## TL;DR - **Projects now delegate instead of executing directly** : you brief the top-level conversation like a team lead, and it decides whether a task needs a new thread or belongs in one already running. - **Each thread is a full cloud session** , complete with its own branch and repo copy, and it keeps running after you close your laptop. - **A retrieval layer (RAG) replaces full-context loading** , so Claude pulls only the relevant chunks of project knowledge instead of stuffing every file into every conversation. - **Team and enterprise plans get permission controls** , letting project owners grant view or edit access by email or organization-wide. - **A library tab collects both uploaded files and Claude’s outputs** , so nothing gets buried across a dozen separate sessions. - **The feature is Claude Code only for now** , cloud-only, limited to Pro and Max subscribers, and still described as an early, rough rollout rather than a finished system. - **A related update, mods, lets developers hook into Claude Code’s harness** , starting with built-in support for reading`agents.md` instruction files alongside or instead of`CLAUDE.md` . ## How does the coordinator delegate work to sub-agents? The core change is architectural. Previously, every chat inside a project was isolated: you’d open a session, run out of context, start a new one, and manually carry decisions across. There was no mechanism for one session to know what another had already decided. In the redesigned version, the project itself is the persistent conversation. When you give it a goal, it determines how to split that work and opens threads to handle each piece. Those threads are not lightweight sub-chats. Each one is a complete Claude Code cloud session with its own branch, and it can further delegate into its own sub-agents and workflows. This mirrors the orchestrator-plus-sub-agent pattern that many developers have already been hand-building on top of Claude: a main agent that routes tasks, and specialized workers underneath that execute them. Because threads run in the cloud rather than in your local terminal, they continue after you close your laptop. That’s a meaningful shift from a single interactive session you have to babysit to something closer to a standing team you can check in on. It also introduces constraints worth knowing: the main coordinating conversation can’t touch connectors like MCPs directly, since those are handled independently inside individual threads, and the whole system is currently cloud-only, meaning it can’t see local files or local tools. ## How does the RAG-based context retrieval work? Every serious Claude user eventually hits the same wall: the project’s context window fills up, and the model starts forgetting earlier decisions or dropping details from uploaded files. The old fix was manual, pruning files, splitting a project into pieces, or starting a fresh session and hoping Claude’s built-in workarounds caught the gap. The new projects add a retrieval layer behind the scenes, effectively retrieval-augmented generation (RAG) applied to project knowledge. Instead of loading every attached file into every conversation, Claude gets a project knowledge search tool that pulls in only the specific pieces relevant to the current question. This runs automatically, with no configuration required. When a project’s total knowledge shrinks back under the context limit, the system reportedly reverts to loading full context again, including files like `CLAUDE.md`, rather than staying in retrieval mode permanently. This is a lighter-weight version of the semantic search and reranking pipelines that some developers have built independently for their own memory systems, but the significant part is that Anthropic has now built a no-setup version directly into the product. It signals a clear direction: context management is moving from something users manually engineer to something the platform handles by default. ## What do the new team permissions actually control? For users on team and enterprise plans, the redesigned projects add sharing and permission controls. A project owner can set who has view-only access versus edit access, and sharing can happen by individual email invite or organization-wide. Someone with view access can see the shared context and decisions made inside a project without being able to act on it directly. - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. This matters for a specific, common problem: two or more people working on related pieces of the same effort, each in their own isolated Claude session, with no clean way to see what the other has already decided. By combining shared project memory with defined permission levels, Anthropic is effectively building a minimal team operating system, one place where a team’s context lives, and one place from which work gets delegated, gated by who’s allowed to see or change what. ## What else shipped alongside the new projects? The redesigned projects beta arrived alongside several related Claude Code updates that point in the same direction, toward more customizable and more autonomous workflows. Claude Code can now read `agents.md`, a project instruction file format also used by other coding agents. By default it’s used when a project has no `CLAUDE.md` of its own; if both exist, `CLAUDE.md` takes priority unless you configure Claude to load both. This support isn’t yet available on Bedrock, Vertex, or Foundry. That feature is also the first public example of “mods,” Anthropic’s emerging system for letting developers hook into the Claude Code harness itself, the software layer that assembles context, calls tools, and renders the session. Mods use function hooks to change behavior such as filtering tool output or altering interface components before the model sees it. Four built-in mods exist in the public repository so far: the agents.md loader, a diff panel, telemetry, and an organization security mod. Custom mods remain early access, and Anthropic notes the underlying API can change between releases. The diff panel mod pairs with a related update: a live panel that shows changed files and line counts updating in real time as Claude edits code or runs commands, viewable in full-screen mode with a sufficiently wide terminal and a git repository present. Two smaller additions round out the update window: `/skill doctor`, which shows the context cost of installed skills so you can disable ones not worth their overhead, and `claude plugin eval`, which lets you run test cases against a plugin (with an automatic comparison run without it) to check whether it actually improves results enough to justify the added complexity. ## Is this a replacement for custom-built agent systems? Not yet, on the evidence available. The new projects setup is explicitly a beta, Claude Code only (not yet in Claude chat or the desktop app, though that’s expected to expand), cloud-only, and limited to select Pro and Max subscribers. Team and enterprise access, along with local execution, are described as coming later. The coordinating conversation also can’t yet reach local tools or MCP connectors directly. More fundamentally, what Anthropic has shipped is a framework, not a finished system. The value people get from custom-built agent setups typically comes from what’s layered on top of that kind of framework: memory systems that distinguish personal from shared data, skills that encode how a specific business actually operates, brand context refined over months rather than a folder of uploaded PDFs, and permission structures that match how a real team is organized rather than a generic default. The new projects give you the scaffolding for that, RAG retrieval, thread delegation, shared memory, permission tiers, but not the accumulated specificity that makes an agent setup genuinely useful for a particular business. ## Frequently Asked Questions ### What is the difference between a Claude Project and a thread? ## Other agents start typing. Remy starts asking. Scoping, trade-offs, edge cases — the real work. Before a line of code. A project is now the persistent, top-level conversation you brief with a goal. A thread is a full Claude Code cloud session that the project spins up to actually execute part of that goal, complete with its own branch and repository copy. ### Does this replace opening multiple Claude Code windows manually? That’s the intent. Instead of manually opening separate windows for separate tasks, you brief the main project conversation, and it decides whether the task belongs in an existing thread or needs a new one. ### Is the new Claude Projects available to everyone? No. It’s currently a beta limited to Claude Code and available to select Pro and Max subscribers. Team, enterprise, and local execution support are planned but not yet available. ### How does the RAG retrieval layer affect existing project files? It changes how Claude accesses them, not what you upload. Rather than loading every file into every conversation, Claude uses a project knowledge search tool to retrieve only relevant sections, which should reduce how often projects hit context limits. ### What are mods in Claude Code? Mods are Anthropic’s emerging system for letting developers customize the Claude Code harness itself, using function hooks to change things like context assembly or tool output. Support for reading `agents.md` instruction files is the first built-in example.