Anthropic rebuilds Claude Projects around parallel Claude Code threads

Anthropic has redesigned Claude Projects, turning what was a folder for storing project context into an orchestration system that hands work to parallel Claude Code sessions. The change was announced on September 17, 2026, according to a dev.to write-up by Nokka that summarizes the official announcement.

A coordinator that delegates to worker threads

Previously, running several sessions against one task meant splitting the work yourself, queuing it, and stitching the results back together. In the new design, you state a goal and a coordinator layer assigns tasks to threads, where each thread is a full Claude Code cloud session.

Each thread operates on its own branch with its own copy of the repository. When two threads touch the same code, the overlap becomes a merge conflict that is resolved like an ordinary pull request. ZDNET, cited in the dev.to post, characterizes this as conflicts being deferred to the end of the pipeline rather than prevented — the system leans on familiar git mechanics instead of new conflict-avoidance machinery.

The announcement offers examples, including one where a web API repository and a mobile repository are connected and the goal is retiring a deprecated v1 endpoint. Claude then creates one thread per repo to move callers, run tests, open pull requests, and report which one must merge first — an implicit grasp of dependency order across repositories.

Threads can also break their assignments into smaller pieces using subagents, loops, and workflows. The dev.to post notes these building blocks are not new — subagents date to April 2026 and dynamic workflows to May 2026 — and that the fresh element is a coordinator layer that persists across sessions.

Shared memory and a library

The quieter change may matter more. Every thread reads from and writes to a shared project memory, which Anthropic says reduces the need for elaborate prompts. The examples given are tacit knowledge rather than code:

• the release day moved to Friday
• why the export feature was cut
• who to consult before touching billing services

None of that information lives in a repository; it exists in people's heads, and the feature is an attempt to externalize it.

Claude also remembers your working and communication patterns, and you can ask it to adjust check-in frequency, how often it opens new threads, and how detailed each update should be. A separate library gathers files you add alongside artifacts Claude produces, so later work builds on earlier output.

A narrow beta with a staged rollout

Availability is narrower than the announcement's headline suggests. The redesigned Projects opened in beta to a subset of Claude Pro and Max members who use cloud sessions in Claude Code and who do not already have Projects on the web or desktop clients. ZDNET's writer reports being excluded by that last condition, since all of their Claude Code work happens in a terminal against local files.

The stated timeline expands to more Pro and Max Claude Code users the following week, then across the rest of Claude plus Team and Enterprise plans, with a waitlist form for those without access. Existing Pro and Max projects keep working unchanged and will be upgraded once the rollout reaches chat and Cowork.

Quota costs and open questions

Anthropic is direct about the trade-off: because each thread is a full Claude Code session, Projects can consume usage limits faster than ordinary work. ZDNET adds a firsthand account of being cut off mid-run after sending a swarm of agents across a codebase. The controls offered are a per-project usage view plus the ability to choose models and effort levels separately for the coordinator and the worker threads.

ZDNET also questions the novelty, arguing that subagent swarms have existed for a while and that Anthropic itself has said the new projects already use this architecture internally — framing Projects as a productized management layer rather than a new capability. The outlet further observed that the feature does not appear fully automatic, with threads seemingly starting only when you press a start button for each one Claude suggests. The dev.to author flags this as the reporter's interpretation, noting the announcement does not confirm it either way.

Threads also run in the cloud only for now; execution on your own machine, beside local tools and code and behind your network, is described as coming soon.

Why it matters

The redesign reframes what a project means inside an AI coding tool: no longer a container for context but an orchestration layer that fans work out and merges it back. Two decisions stand out. Conflicts between parallel threads are handed to standard git processes rather than solved with bespoke machinery, which is legible to anyone who has resolved a pull request. And the shared memory targets the recurring cost of agentic work — re-explaining knowledge that lives in people's heads rather than in repos. The counterweights are real: parallelism is billed in quota, and the eligibility rules mean the developers most likely to run Claude Code locally are also the ones waiting longest.