Anthropic rebuilds Claude Code Projects for parallel agent threads

Anthropic has rebuilt the Projects feature in Claude Code around parallel agent work. Instead of driving one session through a task step by step, a user now describes a goal, and a coordinator splits the work across parallel "threads," each running as its own cloud session. Progress can be tracked either in the main chat or inside each individual thread, including from mobile. Every thread can act on its own: it can open pull requests and run tests without further prompting. Over time, Claude also builds a shared memory across the threads working on a project, and a library collects all the files and results the threads upload or produce, so the work of separate agents stays visible and reusable in one place.

The rebuilt feature is in beta and only partly open: select Pro and Max subscribers who use cloud sessions in Claude Code get access now. Team and Enterprise accounts are told they will get it later, and local execution, meaning threads that run on the user's own machine rather than in the cloud, is described as coming soon. Interested users who are not yet included are directed to a waitlist.

The article frames the move as part of a pattern rather than an isolated update: Anthropic recently made an autopilot mode the default in Claude Code, and the company said that mode outperformed human developers on safety tasks, without the article naming the specific tasks, a benchmark or a study behind that claim. The piece also adds its own editorial reading of the business logic: more autonomous agents running in parallel likely burn more tokens, and doing the work that way moves control over how many tokens get spent from the user to the company selling them, which the article compares to letting a gas company control the throttle in your car. That comparison is the outlet's commentary, not a claim made by Anthropic.

Key facts

Anthropic rebuilt Claude Code's Projects feature so a coordinator splits one described goal across parallel "threads," each running as its own cloud session.
Each thread can open pull requests and run tests on its own, and progress is trackable in the main chat, per thread, or on mobile.
Claude builds shared memory across a project's threads over time, and a library collects all uploaded files and results.
The beta covers select Pro and Max subscribers on cloud sessions now; Team and Enterprise access, plus local execution, are described as coming later, with a waitlist open in the meantime.
The article ties this to Anthropic recently making an autopilot mode the default in Claude Code, which it says Anthropic claims outperformed human developers on safety tasks, though no specific tasks or benchmark are named.

Why it matters

This is a structural change to how Claude Code handles a task, not a new model or a pricing change. Rather than one session working through a goal in sequence, a coordinator now breaks the goal into parallel threads that each run independently in the cloud, act on pull requests and tests without step-by-step supervision, and share memory and a results library across the project. That pushes Claude Code further toward running coding work with less direct human steering at each step, continuing a direction the article says Anthropic already set by making an autopilot mode the default.

Who it affects

Directly, it affects Claude Code users on Pro and Max plans who use cloud sessions, since that is who the beta is open to now. Team and Enterprise users are told they will get access later but do not have it yet. Users who want the feature to run locally rather than in the cloud are told that is coming soon but is not yet available, and anyone not already included can join a waitlist.

How to use it

Access is currently limited to select Pro and Max subscribers already using cloud sessions in Claude Code; the article gives no separate price or plan requirement beyond that. A user works with it by describing a goal to the rebuilt Projects feature, after which the coordinator splits that goal into parallel threads on its own. Team and Enterprise access and local execution are both described only as upcoming, with no date given, and interested users outside the current beta can sign up for the waitlist mentioned in the source.

How solid is it

The description of the feature, the parallel threads, shared memory, the results library, and the current Pro/Max cloud-session beta, is stated directly and specifically. What is not solid is the safety claim invoked alongside it: the article says Anthropic claims its autopilot mode outperformed human developers on safety tasks, but names no specific tasks, no benchmark and no study backing that claim, so it should be read as an unverified vendor claim relayed by the outlet rather than an independently confirmed result.

Risks and caveats

No date is given for when Projects was rebuilt or when autopilot mode became the default, and no version number or release name is attached to either change, which makes both hard to pin down precisely. Team and Enterprise access and local execution are promised only as "later" and "coming soon," with no committed date. The autopilot safety claim rests on an unnamed benchmark. Separately, the article's own point about token spending, that running more autonomous parallel agents likely increases token usage and shifts control over that usage from the user to Anthropic, is the outlet's editorial framing rather than a fact Anthropic has stated, and the source gives no pricing or token-cost figures to test it against.