Claude Code Projects now coordinates parallel agent threads
A coordinator scopes your goal and farms it to cloud sessions on separate branches — shared memory, PR-style merges, but Pro/Max and cloud-only for now.
Copy markdown
One coordinator, many threads
Set a goal plus repo context and Claude “scopes the request, delegates tasks, coordinates the parallel threads, reviews outputs, and assembles the finished result.” Each thread is its own Claude Code cloud session running on its own branch, so a multi-file refactor or a bug backlog gets chewed through in parallel instead of one session at a time.
Shared memory and an artifact library
Threads share persistent project context — release schedules, architectural decisions, even how often you want check-ins and how detailed each update should be. A project library collects the files you add plus artifacts Claude produces, and any thread can split its own work further with subagents, loops, and workflows so big assignments finish faster.
Overlaps resolve as ordinary merge conflicts
When two threads touch the same code, “the overlap is resolved as a merge conflict just like any other PR” — no bespoke locking model to learn. If you already live in branches and pull requests, the mental model is the one you have.
The catch: Pro/Max, cloud-only, and it eats tokens
Beta is limited to select Pro and Max subscribers on cloud sessions with no existing web or desktop projects; broader Pro/Max access rolls out “over the coming week,” with chat, Team, and Enterprise after. Local execution is “coming very soon.” Expect usage limits to arrive faster since every thread burns a full session’s tokens, though you can set the model and effort per coordinator and per worker. No access yet? There’s a waitlist.
Elsewhere: a lab on building self-improving inference
z.ai published a deep-dive on building its own inference infrastructure for “recursive self-improvement” — the stack GLM runs to serve and iterate on itself. It climbed the Hacker News front page today; a grounded look at what running frontier inference in-house actually takes.