跪拜 Guibai
← All articles
Programmer · Artificial Intelligence

Claude Code Projects Gains a Coordinator to Run Parallel AI Agents Without Collisions

By 深蓝AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Parallel AI coding moves from hand-rolled orchestrations to an official feature, so teams lose the burden of building their own scheduling and conflict management, but they gain new quota costs and a cloud-only constraint while the rollout is underway.

Summary

Anthropic has restructured Claude Code Projects around a two-layer orchestration model. A coordinator receives the project goal, breaks it into tasks, and routes each to a dedicated thread. Every thread is a full Claude Code cloud session with its own branch and independent repository copy, so parallel work never collides at the file level. When two threads touch the same code, the overlapping changes surface as a normal pull request merge conflict, handled the same way human teams do it. This turns multi-agent parallelism from a community workaround into an official product feature.

The new model also adds project-level shared memory and an asset library, so facts from one thread become available to all others and new tasks build on past outputs. Costs are real: each thread consumes quota faster, though Anthropic reports coordinator token overhead under 10%. All threads run in the cloud for now, with local execution and private network access promised later. The beta opens to some Claude Pro and Max subscribers who use cloud sessions and have no existing projects, with expansion over the next week and Team/Enterprise later.

Comparing against community DIY setups, the official value is in experience and continuity, not the underlying pattern, which has long existed as orchestrator-worker. The ceiling is the coordinator's decision quality: if task decomposition, escalation, or merge ordering is wrong, the whole system wastes effort. For developers already juggling multiple parallel sessions, the coordinator becomes a need rather than a luxury.

Takeaways
— Projects now has a two-layer structure: a coordinator for direction and threads for execution, each thread running a full Claude Code cloud session with its own branch and repository copy.
— Conflicts between threads are resolved exactly like ordinary PR merge conflicts, reusing git's existing diff and review workflow.
— Threads automatically open PRs, run tests, and suggest merge order; the official example retires a deprecated interface across API, web, and mobile repositories.
— The coordinator's token overhead is under 10% compared to running sessions directly, per Anthropic's Boris on Hacker News.
— All threads currently run in the cloud; local run, private network access, and un-pushed code are not supported yet.
— The beta is limited to some Claude Pro and Max subscribers using cloud sessions and having no existing projects on web or desktop; Team and Enterprise come later.
— Old projects remain unchanged and won't be force-migrated; the two versions coexist for a period.
— Each thread can spawn sub-agents and workflows, forming a three-layer work tree: coordinator -> threads -> sub-agents.
Conclusions

Reusing standard PR merge conflicts for AI-AI collisions keeps the mental model familiar, so developers don't need to learn a new conflict protocol on top of multi-agent orchestration.

The coordinator's decision quality determines the automation ceiling; a flawed task decomposition, escalation, or merge order can nullify the benefits of parallelism.

Making multi-agent orchestration a default product form solidifies the shift from single-session prompting to engineering-team-like management, a direction other coding agent vendors will likely follow.

The cloud-first rollout mirrors the path CI cloud services took, prioritizing standardized workflows, which leaves offline and compliance-sensitive teams waiting for local support.

Concepts & terms
orchestrator-worker pattern
A coordination structure where one central component (the orchestrator) decomposes a goal, delegates tasks to multiple workers, and integrates their results. Claude Code Projects implements it with the coordinator assigning work to threads.
git worktree
A Git feature that lets a single repository maintain multiple working directories, each with its own branch and checkout. It provides manual parallel isolation before a product builds it in.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗