Anthropic has rebuilt the Projects feature in Claude Code around parallel agents. The user describes a goal; a coordinator then splits the work across several threads, each running in its own cloud session. The threads can open pull requests and run tests on their own, and progress is watchable either in the main chat or thread by thread, including from a phone. It is in beta for selected Pro and Max subscribers who already use cloud sessions in Claude Code, by waitlist, with Team and Enterprise access to follow and local execution planned.
Source: the-decoder.com
Over time, Claude builds a shared memory across the threads, and a library accumulates every uploaded file and result the project produces.
The feature follows the direction Anthropic has been pushing for months. The company recently made autopilot mode the default in Claude Code and said it had outperformed human developers on security tasks. Each step moves the same boundary: the developer describes intent earlier and reviews output later, and the span in between that a person actually watches keeps shrinking.
The mechanically interesting part is not the parallelism. It is the coordinator. A single agent working through a task consumes tokens at a rate the user can roughly feel. A coordinator deciding how many threads a goal deserves consumes tokens at a rate set by Anthropic's planning logic, not by the person paying. That is a transfer of control, and it runs in the vendor's favor. The comparison that has been drawn is to a gas company with its hand on the accelerator pedal.
This matters more than usual right now because Anthropic is preparing to go public, and more autonomous agents will almost certainly burn more tokens per unit of work. A product that increases consumption without requiring the customer to ask for more is an unusually clean revenue story to take into an IPO. That is not an accusation of bad faith — parallel agents are a reasonable thing to build — but the incentive and the roadmap point the same way, and it is worth saying so out loud rather than treating the alignment as coincidence.
What the announcement does not cover is the part a working developer would ask about first. There is nothing about what a multi-thread project costs, or any ceiling on how many sessions a coordinator can spin up against one goal. There is nothing about what happens when two threads touch the same files and disagree. And the claim that autopilot beat human developers on security tasks arrives without the benchmark behind it.
A developer who cannot see how many agents are running, or for how long, is in a poor position to argue that the bill was their own decision.