Anthropic's new project coordinator is real. The team-sharing pitch in the video covering it is not, yet.
Anthropic's Sept 17 Claude Code Projects redesign matches its own docs closely. The video calling it a team OS describes sharing that has not shipped.
Contents
On September 17, 2026, Anthropic shipped a real redesign of Claude Code Projects, and a YouTube video covering it the next day gets the mechanics right and leaves the availability unclear in one specific, checkable place. That gap is worth knowing before you plan around this feature, because the part it overclaims is the part a business owner would actually want first: sharing the work with a team.
#What shipped, and what it actually is
The new Projects turns what used to be a folder of context into a coordinating conversation. You describe work in one long-running chat, and Anthropic's own product documentation puts it plainly: "Claude routes each message, and threads do the work and report back." Each thread is a full cloud session on its own branch, running after you close the laptop, and opening a pull request when the work calls for one. Anthropic's launch post frames the same idea under the heading "Threads do the work, Claude directs it."
This is not a co-pilot inside your editor. It is closer to a standing operations desk: you hand it a stream of related work, it decides what becomes a thread and what gets answered in place, and an Overview pane tracks which threads are ready for review, which are waiting on you, and which are still running. A project also keeps its own memory, written to MEMORY.md files the way the docs describe, so a decision you make once reaches every later thread without you repeating yourself.
Two boundaries worth knowing going in, both confirmed directly against the documentation: this only runs in the cloud, so "a local session can't be part of a project" and threads "don't pick up anything from the Claude Code setup on your own machine." And it is scoped to Claude Code specifically, not the older Projects inside claude.ai chat or Cowork, which the docs say "keep working as they do today until the redesigned experience reaches them." The retrieval feature the video also highlights, automatic search over project files once you approach the context limit, is real, but it predates this redesign and already applied to the older Projects too.
#What is not live yet, and why it matters more than the rest
The video frames this as a step toward a team operating system: permission levels, view-versus-edit access, sharing by email, organization-wide rollout. That is where it runs ahead of what actually shipped. Anthropic's documentation is direct about this in its own Limitations section: "A project belongs to one user. You can't share a project or its threads with another user... There are no organization-level controls for projects during the beta." The launch post confirms the timing separately, saying updated projects "across all of Claude and Team and Enterprise plans come after" the current Pro and Max beta.
The video never states outright that sharing is live today, it uses future-tense language like "we'll be able to." But it sits inside a section framed as things Anthropic has "looked at" and "decided to start building," without ever flagging that none of that sharing machinery exists yet, for anyone, during this beta. A viewer skimming the video would reasonably walk away thinking a team could split this work up right now. As of this week, per Anthropic's own product docs and launch post, you can't.
#What an owner-operator should actually do with this
The real, defensible story here is that a frontier lab just shipped, natively, a pattern a growing number of teams have been hand-building: one coordinator, a fleet of parallel worker agents, shared memory, and automatic retrieval over what the project already knows. It rhymes with the shape we have argued knowledge agents need: work decomposed into orchestrated sub-tasks, with a verification step before any result is trusted. That validates the architecture. It does not make it a differentiator by itself anymore.
What still separates a generic setup from one that actually runs your business is depth the pattern alone doesn't provide: skills that encode how your specific operation works, context built up over months rather than a folder of uploaded PDFs, and a permission model that maps to how your team is actually structured, not one user's single account. That last piece is precisely the part not shipped yet. If your plan for this quarter depends on splitting reviewed work across a team inside one project, on Team or Enterprise, that plan is not buildable today. It is worth testing solo, on Pro or Max, to see whether the coordinator-and-threads pattern fits how you already hand off work, and worth checking Anthropic's own documentation again before you commit anything larger, since this is a beta under active, near-term rollout.
#Sources
A new era.
Room for you.
Keep reading
- · 9 min
A 40-year-old search algorithm beats modern embeddings at agentic research. Only one version of it does.
A Hornet CEO says BM25 is unreasonably effective for agentic search. The paper he cites disagrees, then a newer paper agrees. The condition between them is the real finding.
- · 6 min
Chunking is not dead. The talk that says so smooths one of its own numbers.
An AI21 researcher attacks query-dependent chunk size with a 17-year-old fusion trick. The mechanism holds. A separate 2026 benchmark found the opposite on a different corpus.
- · 5 min
"Discovery intelligence" is a real research term. The video's argument with MIT about it is built on a claim MIT never made.
A new paper's numbers on self-improving scientific agents check out under verification. The video covering it disagrees with MIT by inverting what MIT actually says.