What Claude Code Projects' parallel "streams" are
The core of this update is the stream: an independent workflow inside a Project, where each stream is effectively a cloud Claude Code session with its own branch and its own copy of the repository. Instead of working one task at a time, a team can launch several streams at once — one for the backend, one for the frontend, one for tests — and a central coordinator assigns tasks, tracks progress on each, and flags conflicts between them before they ever reach the point of merging.
It's the same problem this site already solves at a smaller scale with specialized subagents (audit, writing, HTML building) — except Anthropic is now formalizing it as a native platform feature, with real branch isolation instead of manual coordination between sessions.
Shared memory: agent B doesn't repeat what agent A already found
The classic problem with running AI agents in parallel is that each one starts from zero, with no idea what the others have already learned. Claude Code Projects solves it with a shared memory layer across the agents in the same project: goals, file libraries, artifacts and history stay visible to every active stream. If stream A finds and documents a bug in a shared artifact, stream B can just read it instead of tripping over the same problem and wasting time diagnosing it again.
This is exactly the kind of friction that shows up whenever work gets delegated to several agents without shared context — and it's the reason this project prioritizes keeping a shared log (memory, session notes) instead of repeating research that's already been done.
Pull-request-style merging: human control doesn't disappear
Streams don't merge on their own. When two agents touch the same code, the system generates merge conflicts that the developer reviews and approves manually, following the same pattern as a traditional Git pull request. It's a meaningful design choice: Anthropic isn't automating the final decision about which code change survives — only the coordination that happens before that point, to get there faster.
On the security side, each stream runs in an isolated sandbox with no network access by default, and it's the organization's admins who configure which domains each agent can reach. There's also a per-stream context limit, meant to keep each agent focused on its own task instead of losing quality by trying to cover too much at once.
What this means if you build software (or content) with AI
For a small team or a freelancer who already uses Claude Code as part of their workflow, this update doesn't change the "what" — you're still writing software with AI's help — it changes "how much at once": it lets you split a real project (say, the backend, frontend and tests of a custom system) into parallel streams without losing track of who did what, and without having to manually juggle half a dozen open sessions just to keep the thread.
The use case Anthropic's own announcement describes — simultaneous work on backend, frontend and tests within one project — is the same pattern that already applies to custom software and AI-integration projects for businesses: more room to parallelize technical work without giving up human control over which change gets approved.
Talk about your project →