Claude Code rebuilds Projects around a coordinator for parallel cloud agents, and speed now draws down your usage budget
Anthropic relaunched Projects in Claude Code on September 17 as a coordinator that splits a goal across parallel cloud agents, each on its own branch with shared memory. The threads keep running after you close the laptop, and they draw down usage limits faster.
Anthropic rebuilt Projects inside Claude Code on September 17, and the change is larger than a UI refresh. Instead of a folder that holds context for a single chat, a Project is now a coordinator: you describe an engineering goal, and Claude scopes it, splits the work across parallel cloud sessions, reviews the output, and assembles the result. The pitch is that you stop babysitting several separate agent threads like browser tabs nobody wants to close. The trade is that running those threads at once spends your plan faster, and more of the throttle now sits with Anthropic.
From a folder to a coordinator
The old Projects were a container: a place to park files and instructions so a single conversation had the right context. The redesign turns that container into an orchestrator. Each unit of work is a “thread,” which is a Claude Code cloud session running on its own branch and its own copy of the repository. Because the sessions run on Anthropic’s infrastructure rather than your machine, the work continues after you close the laptop, and you can steer it from a phone. The main project chat becomes a control surface where you set the goal and check progress, while the individual threads do the hands-on work.
How the parallel threads work
The coordinator handles the parts a developer usually does by hand across multiple sessions: it delegates tasks, tracks what each thread is doing, and stitches the outputs back together. When two threads touch the same code, the overlap surfaces as a merge conflict, resolved the same way as any other pull request, which keeps the model familiar rather than inventing a new reconciliation scheme. Threads also read from and write to a shared memory, so a project can retain decisions like a shipped release date or which service owner to check with before touching billing, without re-explaining it in every prompt. A project library collects both user-added files and Claude-generated artifacts.
Anthropic’s own examples show the intended shape. Set a goal to cut checkout p75 latency, and Claude profiles each endpoint, tests optimizations, and opens pull requests across parallel threads. Point it at API, web, and mobile repositories with a goal of retiring a deprecated v1 endpoint, and it spins up a thread per repository to migrate callers, run tests, and open PRs, then reports which need to merge first. Both are coordination-heavy jobs that normally eat a developer’s attention rather than their skill.
The bill comes faster
The convenience has a metered cost. Anthropic notes directly that Projects can reach usage limits faster because multiple threads run simultaneously, and it points users to model selection and effort levels as the levers to manage that. Independent coverage read the incentive the same way: The Decoder observed that more autonomous agents tend to mean more tokens burned, and that control over how many get burned shifts from the user toward the company. That is not a verdict on whether the feature is worth it, but it is a real change in who holds the throttle, and it belongs in the cost math before a team leans on parallel threads for routine work.
Access is narrow at launch. The beta opened September 17 to a select group of Claude Pro and Max subscribers who already use cloud sessions in Claude Code and have no existing projects on the web or desktop, widening to more users on those plans over the following week. Team and Enterprise plans come later, so this is an evaluation surface for individual developers first, not something to build a team workflow on yet.
What it means for developers
For a developer facing a multi-repo migration or a performance sweep, the redesign collapses the coordination overhead that makes those jobs tedious: branching, tracking, and reassembling parallel work now sit with the coordinator. The merge-conflict model keeps the mental overhead low, since the resolution path is the one every team already runs. The counterweight is control. You trade hands-on-keyboard steering for orchestration you see less of, and you accept a usage curve that climbs faster than a single interactive session. For read-heavy, well-scoped work the trade looks favorable. For ambiguous changes where a wrong turn is expensive, the reduced visibility is the part to weigh.
What to watch
The broader pattern is that model vendors are competing to own the agent runtime, not just the model. OpenAI put the Codex harness behind a single Agents API call earlier this month, and Anthropic is making the same bet one layer up, at the project and workflow level. The questions worth tracking over the next quarter are whether the coordinator holds up on messy real-world changes rather than clean demos, whether the parallel usage cost becomes predictable enough to plan around, and how the experience lands when it reaches Team and Enterprise plans where the coordination pain is largest. Stackmaven’s follow-up will revisit around December 17.