Claude Code Projects

Canonical version: Claude Code Projects.

Projects is a Claude Code feature launched in beta on 2026-09-17, on desktop and web (cloud sessions), for select Pro and Max users. Anthropic says it's coming to all Claude users soon.

A project is one conversation with Claude. You describe what needs doing, and Claude splits the work into threads by itself, runs them as parallel cloud sessions, passes context between them, and keeps going after you close your laptop.

Not to be confused with the older Projects feature of the Claude app, which is a shared knowledge base for chats (see Claude).

How it works

• One orchestrator agent per project. It holds the memory and spins off subagents for each task (AI Subagents)
• Each thread runs as its own cloud session (Claude Code Web)
• Context flows between threads through the orchestrator, so you don't copy-paste between sessions
• You can ask it to be proactive or to do things on a schedule
• It mixes expensive and cheap models depending on your preferences

Thariq Shihipar (Claude Code team) describes it as bringing the architecture of Claude Tag to Claude Code: one long-lived agent that manages memory and delegates, instead of a pile of independent sessions.

Why it matters

Until now, running Claude Code in parallel meant I was the orchestrator. Open five sessions, remember which one does what, shuttle context between them by hand. Claude Code Agent View made that less painful, but I was still the one managing sessions.

Projects moves the orchestration into the product. Boris Cherny put it this way: he stopped managing sessions, sends thoughts as they come, and lets Claude split them into threads. The project remembers how he works. That's the shift: from managing agents to talking to one agent that manages the others.

Ethan Mollick's test shows how far it stretches. He asked Claude Fable 5.1 in a project to pick famous historical mysteries and try to resolve them. It launched research agents, picked the mysteries, then spun up eighteen threads (one per mystery). Each thread launched more agents (avalanche simulations, code breaking), and a final set of skeptical agents fact-checked the write-ups. About a day of work, with the orchestrator coordinating the whole thing. In his words, it basically creates an organization to solve your issue.

My take: this is the same direction as AI Agent Orchestration everywhere, but packaged for people who don't want to write orchestration code. The open questions are cost visibility (a project can fan out to dozens of agents) and how much you can audit what each thread did.

Compared to other Claude Code multi-agent features

• Claude Code Agent Teams: local, experimental coordination between Claude Code instances
• Claude Code Agent View: one screen to dispatch and watch background sessions; you're still the orchestrator
• Claude Code Routines: saved cloud sessions with triggers; one task, not a living project
• Dynamic workflows (ultracode): deterministic, script-driven fan-out inside a single session
• Projects: a persistent conversation that plans, delegates and remembers across threads

References

• Announcement (ClaudeDevs, 2026-09-17): https://x.com/ClaudeDevs/status/2100633571543367691
• Thariq on the Claude Tag architecture: https://x.com/trq212/status/2100637478453686593
• Boris Cherny on how it changed his coding: https://x.com/bcherny/status/2100669598995816511
• Ethan Mollick's historical mysteries experiment: https://x.com/emollick/status/2100665801800093959
• Mollick's results: https://historical-mysteries.netlify.app

Related

• Claude Code
• Claude Tag
• Claude Code Web
• Claude Code Agent View
• Claude Code Agent Teams
• AI Subagents
• AI Agent Orchestration

About Sébastien

Ready to get to the next level?