For years, “Projects” in AI products meant a folder with a personality. Dump files in. Paste instructions. Hope the next chat remembers why billing is sacred. That was storage cosplay.
On September 17, 2026, Anthropic relaunched Projects inside Claude Code as a multi-agent runtime. You set a goal. A coordinator scopes work, spins parallel threads, reviews outputs, and assembles results. Each thread is a full Claude Code cloud session on its own branch and repo copy. Collisions become merge conflicts like any other PR. Shared memory and a project library carry decisions and artifacts across the week.
This review is primary-source. I did not run a signed-in hands-on pass on the beta. What follows is Anthropic’s blog plus corroboration from The Verge, Reworked, and Superpower Daily, and what marketing, GTM, and eng-adjacent operators should do with a swarm that bills like sessions and fails like git.
What it is
Anthropic’s Claude blog post (“Projects redesigned: from folder to conversation”) is the primary. Updated projects are available in beta in Claude Code. The pitch: stop dividing work, juggling handoffs, and stitching results yourself. Describe what needs to get done. Claude manages the work.
Mechanically, a project has threads that do the work and a coordinator that directs them. You pick a goal and the repo or context. Claude suggests work it can pick up. You configure the cloud environment, connectors, plugins, instructions, and model. You monitor in the main project chat or dive into a thread. Anthropic’s briefing metaphor is chief of staff: brief the project, and it routes to new or existing threads.
Under the hood, each thread is a Claude Code cloud session working on its own branch and copy of the repo. The coordinator keeps work organized. Overlap on the same code resolves as a merge conflict “just like any other PR.” Threads can further split work with subagents, loops, and workflows.
Context is the second half of the product. Projects target long-running or agentic workflows: longer than one reply, more than one part. Every thread adds to and draws from shared memory. Anthropic’s examples: the release moved to Friday; why an export was dropped; who to check before touching billing. A library collects files you add and artifacts Claude produces. Claude also remembers working and communication style: check-in frequency, how often to open threads, update detail.
The Verge (Stevie Bonifield, September 17) frames the same shape for a general audience: a team of Claude Code agents under one roof with shared memory, goals, and a library of files and artifacts. Threads run tasks in parallel. A coordinator directs. Users interact with threads individually or through the main project chat. At launch, threads run in the cloud; local tools and code are “coming very soon.”
What changed
Three shifts matter more than the UI chrome.
First, Projects become a control plane. Old model: chat with files attached. New model: goal with workers. Anthropic is packaging the multi-session leap as the Projects surface people already know.
Second, isolation is honest about concurrency. Parallel agents without branches become silent overwrite theater. Each worker stays on a branch; collisions are merge conflicts. That is not a bug. It is the product admitting parallel authorship still needs a human merge ritual.
Third, continuity without a laptop. Anthropic says you can steer from your phone and work continues after you step away. Reworked (Siobhan Fagan, September 17) repeats the laptop-closed loop. Shipping cadence changes: a migration can burn tokens overnight while a PM reviews from a commute.
Access is gated. Anthropic: beta for select Pro and Max subscribers on Claude Code cloud sessions with no existing web or desktop projects. Over the coming week, more Claude Code users on those plans. Updated projects across all of Claude, plus Team and Enterprise, come later. Waitlist if you lack access. Existing projects keep working; upgrades roll as chat and Cowork expand. The Verge adds later expansion to all Pro, Max, Team, and Enterprise, plus Cowork and regular chats.
Mechanism (what you are actually buying)
Strip the demo. You are buying four coupled pieces: a coordinator that scopes, delegates, reviews, and assembles; worker threads that are full Claude Code cloud sessions; shared memory plus a library so week-long work does not restart each morning; and a collision policy that surfaces conflict as git-shaped truth.
Anthropic’s examples show the scale. Reduce checkout p75 latency: profile endpoints, test optimizations, open PRs in parallel. Retire a deprecated v1 across API, web, and mobile: one thread per repo, migrate, test, open PRs, report merge order. That is program management for agent labor, not “write a function.”
Superpower Daily (September 17) names the week-one trade: coordination convenience versus cloud-only workers and faster usage burn. No separate Projects SKU. Coordinator and workers draw the existing Claude allowance. Anthropic warns projects can hit limits faster because each thread is a full session. Check project-specific usage; set model and effort for coordinator and workers.
Pricing context (Anthropic plans page, accessed September 18): Pro is $20 monthly or $17 monthly annual ($200 up front), including Claude Code and Projects. Max starts from $100 monthly (5x or 20x Pro usage) with early access to advanced features. Claude Code on paid plans shares the chat usage pool. Team and Enterprise list Claude Code and Projects, with sharing on those plans. Do not confuse the pricing-table “Projects” line with redesigned Claude Code Projects beta gates. The blog’s access rules control this week.
What works (on paper)
Parallel work with a real merge story. If your team lives in PRs, refusing to pretend two agents edited the same file cleanly beats a magic auto-merge claim. The honesty layer is the sell for operators burned by agent overwrite.
Goal-level briefing beats prompt spaghetti. Brief a coordinator like a chief of staff instead of pasting constraints into five chats. Shared memory compounds: day two of a migration costs less context engineering than day one.
Cross-repo vignettes match how agencies and product teams ship. API plus web plus mobile is the default cut. Threads per repo plus merge-order advice is delivery management, not autocomplete.
Mobile steerability plus cloud continuity is a management surface. Assign a latency goal at 6pm, check threads from a phone. That is who watches the factory floor.
Model and effort dials per coordinator versus workers are the cost controls. Swarm without dials is how finance finds you.
What breaks (or at least what to distrust)
Cloud-only is a hard wall. Anthropic says local tools, local code, and behind-network work are coming “very soon,” with no date. If your monorepo cannot leave the network, this beta demos the control plane; it is not your production harness.
Usage math will surprise people who think one project equals one chat. Each thread is a full session. Three workers plus a chatty coordinator can exhaust Pro or Max faster than one deep coding night. Check project-specific usage before you greenlight a swarm.
Merge conflicts do not disappear. Superpower Daily is blunt: developers still resolve overlapping edits. If your AI story was “agents merge themselves,” this product disappoints on purpose. Budget human merge time, or you will misread the honesty layer as a bug.
First-wave beta excludes people with existing web or desktop projects. Awkward for power users already living in Projects. Waitlist and staged upgrade mean “we all have Projects” is not one product surface this week.
This review did not verify thread reliability, PR quality, tests, memory, or phone UX hands-on. Coverage matches Anthropic on architecture and access. Pilot the trust breakers: bad merge-order advice, stale memory, stuck workers burning tokens.
Who it is for (and who should wait)
Best fit this week: Pro or Max users on Claude Code cloud sessions, without blocking existing projects, running multi-part work (migrations, performance passes, multi-repo cleanups) with a PR culture that can absorb agent branches.
Also fit: eng-adjacent GTM and marketing tech teams that ship internal tools, landing experiments, or agency deliverables as real repos. If a “campaign build” is three services and a design system, a coordinator beats a chat tab.
Wait if you need local or private-network access now; you are on Team or Enterprise waiting for later rollout; you wanted a prettier folder, not a swarm bill; or you cannot staff merge review for agent PRs.
Marketers who never touch a repo still feel this. Swarms change agency SLAs and internal tool cadence. Overnight parallel PRs turn “build by Monday” into a negotiation about review capacity, not coding capacity. That is GTM ops in an engineering costume.
What operators get wrong
They will treat Projects as a file tree with vibes, under-brief the coordinator, and over-trust the first assembled result. The product is the runtime. Brief goals, constraints, merge owners, and check-ins like a contractor team.
They will max threads because parallelism looks productive, then hit limits mid-migration. Start with two threads and a hard usage ceiling. Expand only after clean PRs.
They will skip naming a human merge owner. Conflicts are honesty only if someone owns the resolve. Put that name in project instructions on day one.
They will pilot on sacred billing. Anthropic’s own memory vignette implies caution. Start with a latency goal or a non-critical deprecated endpoint.
They will conflate this with the September 16 One Claude chat/Cowork merge and Docs/Slides story. Different surface. One Claude is the docs/decks workspace. Claude Code Projects is the multi-agent coding runtime. Do not brief leadership as one SKU move.
What to do this week
- Check access. Pro or Max on Claude Code cloud sessions with no existing web/desktop projects: look for the beta. Otherwise join the waitlist and calendar the coming-week expansion for more Claude Code users on those plans.
- Run one bounded pilot. One performance goal on a non-sacred path, or one deprecated endpoint across at most two repos. Cap threads. Set model and effort for coordinator versus workers. Watch project-specific usage after two hours.
- Write the merge protocol before the swarm. Who resolves conflicts? Who merges first when the coordinator suggests an order? What tests stay green? Put it in project instructions so memory inherits process.
- Marketing and agency leads: update delivery language. “Build by Friday” now depends on overnight agent capacity plus human PR review. Add review hours to SOWs. Stop selling coding hours alone.
- GTM operators owning internal tools: inventory multi-repo, conflict-prone roadmap items. Those are Projects candidates. Single-file copy changes are not.
- Do not promise Team or Enterprise org-wide enablement from the September 17 blog alone. Sequence is Claude Code Pro/Max first, then broader Claude, Team, and Enterprise. Match change management to that.
Sharp close
Claude Code Projects is Anthropic admitting the unit of work is no longer a chat. The unit is a coordinated swarm with shared state, and the invoice is session-shaped. The merge conflict is not a failure mode they forgot to polish away. It is the product telling you parallel agents still live in the same universe as human PRs.
Buy the prettier-folder story and you under-configure the coordinator and overspend on threads. Buy the runtime story and you brief goals, cap usage, name a merge owner, and supervise overnight progress like a factory from your phone.
Operator decision this week: organizing files, or running a small agent company inside one project?
