Anthropic redesigned Claude Projects on September 17, and the framing in its own post is exactly right: “from folder to conversation.” The old Projects was a container — you dropped files and chats into it and organized your own context by hand. The new one is a single, persistent conversation that acts as a coordinator: you describe what needs to get done, and Claude scopes the request, delegates the work across parallel threads, reviews what comes back, and assembles the finished result. It’s in beta today for a narrow slice of Claude Code’s cloud-session users, and it’s the clearest sign yet that Anthropic sees “Projects” not as a filing cabinet but as a standing engineering team you talk to.
How the coordinator actually works#
Strip the product-page language and the mechanism is concrete. Each Project runs a coordinator thread that directs work; the actual work happens in separate threads, each of which is an independent Claude Code cloud session running on its own repository branch. Those threads aren’t limited to single tasks — Anthropic’s own description says threads “can create subagents, loops, and workflows for complex tasks,” meaning a single Project can fan out into the same Agent Teams and Dynamic Workflows machinery this blog has covered since March and tracked through the June architecture roundup. Overlapping changes from different threads resolve the ordinary way software already knows how to resolve them: as git merge conflicts, reviewed like any other.
The other real addition is memory that spans the whole Project rather than a single chat. Anthropic’s example is mundane and telling: Claude can now retain that a release moved to Friday, why an export got dropped, or who needs to sign off before someone touches the billing service — details that used to evaporate the moment a session ended, or that had to be re-explained in every new chat. That’s a direct answer to the kind of context loss that made long-running, multi-week engineering work with any chat-based tool exhausting to steer.
What’s actually shipping today, and what isn’t#
The beta is narrow by design. It’s available to “select Claude Pro and Max subscribers who use cloud sessions in Claude Code and don’t have any existing projects on the web or desktop” — a carve-out that conveniently avoids migrating anyone’s existing folder-based setup on day one. Anthropic says access widens to more Pro/Max users “over the coming week,” with Team and Enterprise, plus integration into chat and Cowork, coming later.
The bigger limitation is architectural, not just a rollout gate: threads run in the cloud only. Anthropic’s own post says local-machine execution — “running on your machine alongside your local tools and code and behind your network” — is “coming very soon,” which is a polite way of saying it isn’t here yet. That’s a real gap for exactly the teams most likely to want this: anyone whose build depends on a VPN-gated internal service, a local database, or credentials that were never meant to leave a laptop can’t use the coordinator model at all until that lands.
The part worth taking seriously: this can drain a plan fast#
Anthropic doesn’t bury this, and neither should coverage of it: because each thread is a full, independent Claude Code cloud session, running several of them in parallel means running several full sessions in parallel. The New Stack flagged this directly hours after launch — a Project that fans out into four or five concurrent threads to parallelize a big refactor can hit a Pro or Max plan’s usage limit well before a single-threaded session would have. Anthropic’s mitigation is visibility rather than a cost cap: per-project usage monitoring, plus the ability to choose the model and effort level independently for the coordinator thread and the worker threads, so a team can run a cheap, low-effort coordinator directing a smaller number of high-effort workers instead of everything running at maximum cost by default. That’s a real lever, but it puts the burden of not over-fanning-out squarely on the user, in a beta most people will be trying out for the first time this week without much intuition for where that line is.
The convergence worth noting#
There’s an interesting coincidence in timing here, and it’s worth being precise about what it is and isn’t. One week ago, Cursor launched its own “Projects” — a coordinator agent that directs worker agents, subscribes to Slack and PR events, and runs cloud-resident after the laptop closes. Now Anthropic has redesigned its own, differently-scoped, pre-existing “Projects” feature around the same coordinator-plus-parallel-threads shape. Seven days is too short a window for one company to have built a reactive clone of the other — what’s actually happening is that both companies are exposing the same underlying idea (a standing coordinator that delegates to a pool of workers) through a product surface that happens to share a name. The difference is what’s underneath. Cursor’s coordinator is new infrastructure bolted onto an IDE-anchored product that spent three years assuming a human watches one chat panel at a time. Anthropic’s is a conversational front end for orchestration primitives — Routines, Agent Teams, Dynamic Workflows — that have been running in production, cloud-native and headless, since April. This launch isn’t Anthropic inventing multi-agent coordination; it’s Anthropic finally giving people who don’t want to learn /agents syntax a plain-language way to drive the system that was already there.
That’s a genuine product improvement, and the honest caveats — cloud-only, narrow beta, and a usage model that can surprise you — are exactly the kind of rollout friction this blog has flagged in Anthropic’s own past missteps, from the silent effort-default change to the three-bug postmortem. Anthropic disclosing the cost tradeoff itself, on day one, in its own launch post, rather than letting a third party discover it, is the difference that actually matters.
Sources: Anthropic — “Projects redesigned: from folder to conversation” (Sept 17, 2026, primary, direct fetch); The New Stack — “Anthropic’s new Claude Code feature could drain your plan before lunch” (Sept 17, 2026); this blog’s prior coverage of Cursor Projects and multi-agent architecture patterns.