Anthropic rebuilds Claude Code Projects around parallel cloud agents

Anthropic has finished a full rebuild of Projects in Claude Code, giving the feature native multi-agent support. A developer now states one high-level goal in the main conversation — for example, refactoring an entire authentication module to OAuth 2.0 and filling in automated tests — and Claude spawns multiple Git branches in the cloud to work in parallel, opening pull requests when the work is done. Projects is no longer just a folder for prompts and reference documents.

At the center of the new architecture is a Coordinator. It decomposes a complex engineering task into subtasks and dispatches them to parallel work threads. Each thread is a Claude Code instance running in its own isolated cloud environment, with an isolated copy of the code and its own Git branch, writing code and unit tests for different modules at the same time. When a thread finishes, it commits in the background and generates a pull request, while the Coordinator tracks the context dependencies between subtasks and tells the developer which branch to merge first and which module to test afterwards.

To address the old problem of parallel agents each working in isolation and losing context, the update adds shared memory. Technical conventions, architectural decisions and team preferences that accumulate inside each branch thread sync back to shared memory in real time, so Claude can remember that a release date moved to Friday, why an export feature was cancelled, or whom to contact before touching the billing service. The process is fully asynchronous: a developer can issue an instruction, close the laptop and walk away, or check thread progress and fine-tune instructions from mobile.

Anthropic concedes that multi-agent work is not cheap, and advises users to watch their usage. The redesigned Projects is in beta, initially limited to some Claude Pro and Max subscribers who use cloud session projects; access will widen over the coming week, while Team, Enterprise and Cowork users still have to wait.

Alongside the product rebuild, Anthropic published a report titled Measuring the Pace of AI Development that discloses its internal benchmarks for automated AI research for the first time. Under the Epoch AI autonomy scale the company uses — AL0 for pure human work, AL3 for deep human-machine collaboration, AL4 for an AI completing a task end to end from a single high-level instruction, and AL5 for a fully autonomous loop with no human intervention — 26% of Anthropic's R&D tasks currently reach AL4, up from less than 1% in February this year.

That share rests on a very large internal agent fleet. As of August, the company's internal research platform was running as many as 30,000 AI agents concurrently, writing code, running experiments and reshaping architecture in the cloud, triggering more than one billion decisions in a single month. Lower the bar to AL3 and more than 90% of Anthropic's R&D workflow already sits in the human-machine collaboration band.

Asked what happens if tens of thousands of agents run wild, Anthropic points to Agent Oversight, an around-the-clock system that reviews AI operations. Out of more than one billion model decisions in August, it says the system intercepted roughly 0.002% as high-risk or anomalous. The report also produces a receipt for safety spending: between July 13 and 20, about 6% of the company's AI R&D compute went to safety and alignment research, while roughly 12% of AI-driven R&D compute went to safety-related tasks in the same week. Anthropic adds that it welcomes third-party evaluators to audit its automation data with the same level of access granted to its internal risk team.

Anthropic is not alone on this path. On September 10, Cursor shipped Projects in beta, also built around a Coordinator agent — a coordinator that writes no code itself but dispatches thousands of sub-agents, with project context shared in the cloud so work continues after the laptop is closed. The implication is clear: writing code keeps losing weight in the engineering chain, while decomposition and scheduling become the new core skill. Multi-agent parallelism is no cure-all, though. Overlapping edits across branches still collide in Git merge conflicts, and when that happens the Coordinator can only untangle the dependencies — a human engineer still has to step in and decide.