Anthropic has relaunched Projects inside Claude Code as a command center that runs several AI agents at once in the cloud, The Verge reports. Instead of driving one session step by step, you state a high-level goal and a coordinator agent breaks it into subtasks and dispatches them to parallel work threads.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Claude Code Projects Now Runs Multiple AI Agents in the Cloud |
| Short answer | Anthropic rebuilt Projects in Claude Code as a cloud command center: one coordinator splits a goal into parallel work threads, each an isolated Claude Code instance with its own Git branch. |
| Reading time | ~2 min |
| Last updated | 2026-09-18 |
How the new Projects works
- A coordinator sits on top. You hand it one high-level goal - The Verge’s example is restructuring a whole authentication module - and it decomposes the work itself.
- Each work thread is a separate Claude Code instance in its own isolated cloud environment, with its own copy of the code and its own Git branch.
- Threads commit and open pull requests in the background, and the coordinator tracks dependencies between subtasks so you know what to merge or test first.
- A shared memory carries technical standards, architecture decisions and team preferences back into the project, so you stop re-pasting the same setup context into every thread. QbitAI walks through the same architecture in Chinese.
Asynchronous by design - and priced like it
The flow is asynchronous: Anthropic’s pitch is that you can hand over a goal, close the laptop, and check thread progress from your phone. The company also says plainly that multi-agent runs are not cheap, and QbitAI’s coverage repeats the advice to watch your usage while several agents run at once.
Projects is in beta and gated. The first wave is a subset of Claude Pro and Max subscribers who use cloud-session projects, with Team, Enterprise and Cowork to follow over the next week.
The numbers Anthropic published alongside it
Anthropic’s “Measuring AI R&D pace” report, as summarized by QbitAI, gives a sense of how far it has already pushed this internally:
- 26% of internal R&D tasks now sit at AL4 - an AI completing an end-to-end task from one high-level instruction - up from under 1% in February.
- More than 90% of its R&D workflow is at AL3, deep human-AI collaboration.
- Roughly 30,000 agents run concurrently on its internal platform as of August, producing more than a billion model decisions a month; its oversight system flagged about 0.002% of them as high-risk or abnormal.
What this means if you run an agent on your own machine
Read the two halves of this story together. Cloud orchestration is getting better at producing artifacts - branches, pull requests, documents. What it does not change is where the work lands and who approves it.
If the work is your own files, folders and desktop apps, the question that matters is the one a local client answers: does the agent plan first, ask before touching anything, and leave a trail you can audit afterwards? MOOGH is built for that side of the split - it runs on your Windows machine, shows you the plan, waits for approval, and logs what it did. Download MOOGH if you want the approving to stay with you.
Related reading: