刚刚，Claude Code大重构！内部3万Agent管理技术免费开放

TL;DR - Anthropic redesigned Claude Code Projects around a coordinator that decomposes development goals across parallel, cloud-hosted agents with isolated Git branches, shared memory, and automatic pull requests. The beta reflects a broader shift from AI code completion toward asynchronous, multi-agent software engineering.

• Each worker runs an independent Claude Code instance, writes and tests code on its own branch, and submits a PR while the coordinator tracks dependencies and merge order.
• Shared memory synchronizes technical conventions, architecture decisions, and team preferences across agents to reduce context loss and repeated prompting.
• Anthropic reports operating up to 30,000 internal agents concurrently, with 26% of its R&D tasks reaching the cited AL4 level of end-to-end execution from high-level instructions.
• Projects is initially a limited beta for selected Claude Pro and Max users; Anthropic warns that parallel agents can significantly increase usage costs.