Claude Code gets a chief of staff for your repos

Managing multiple AI coding sessions usually turns you into an exhausted manager. You end up copying context between tabs, tracking branches manually, and stitching disconnected PRs back together. Anthropic wants to eliminate that headache entirely.

Anthropic has rolled out a major beta redesign of Projects inside Claude Code. Instead of serving as static folders for instructions and files, Projects now act as agentic coordinators capable of running and orchestrating parallel engineering threads.

The beta is live for select Claude Pro and Max subscribers who use cloud sessions, with broader access and team tiers rolling out soon.

A chief of staff for your terminal

The redesigned structure splits work into two distinct roles: a central coordinator and worker threads.

You brief the main Claude chat just like you would brief an engineering chief of staff. You define the high-level objective, and Claude scopes the work, creates parallel threads, assigns tasks, reviews outputs, and packages the result.

Anthropic gave the example of cutting checkout latency. You set the goal, and Claude spins up threads to profile individual endpoints, run benchmarks, and open separate PRs at the same time. If you need to deprecate an API, Claude can connect across your web, mobile, and backend repositories simultaneously, migrating callers in each repo and telling you which PR to merge first.

Under the hood, each thread runs as an isolated Claude Code cloud session with its own branch and repository clone. If two threads touch overlapping code, it gets resolved as a standard Git merge conflict. Threads can even spawn their own subagents and loops for heavy tasks.

You can steer the whole operation from your phone or walk away entirely while the agents keep grinding in the cloud.

Shared memory, shared baggage

One big hurdle with multi-agent setups is context drift. Anthropic tackles this with a shared memory system and project library across all threads.

Claude logs architecture decisions, operational constraints, and workflow quirks along the way. If a release moved to Friday or a specific teammate must sign off on billing changes, every worker thread picks up that context without requiring prompt gymnastics.

The catch is token consumption. Because every parallel thread is a full Claude Code session, running multiple streams will burn through your rate limits in a hurry. Claude lets you configure effort levels and specific models for both the coordinator and worker threads to keep usage in check.

Right now, threads only run in cloud environments, though Anthropic says local machine support is coming very soon.

Why it matters

We think Anthropic made a brilliant design choice by anchoring agent coordination to standard Git workflows. Instead of inventing proprietary orchestration layers, it treats agents like real junior developers: working in separate branches, opening PRs, and resolving conflicts through standard version control.

For solo developers and small teams, this shifts AI assistance from inline code completion to genuine delegation. Tedious cross-repo migrations and performance profiling can finally happen asynchronously while you focus on core architecture.

Still, whether you should rely on it today depends on your workflow. Cloud-only execution is a dealbreaker for setups behind private enterprise VPNs or complex local dev containers, and aggressive rate limits will frustrate heavy users. But by turning Projects into an agent coordinator, Anthropic just pushed the coding assistant battleground far beyond the IDE chat window.