Projects is back inside Claude Code, and this time it is built to run several agents in one workspace. They share memory, a goal list, and a file library. Each project can spin up "threads" on different tasks at the same time. A "coordinator" sits on top and directs them.
Keeping that group under one roof, instead of a mess of separate chats, is the pitch. The Verge's account also flags the resemblance to Grok Bot and other tools that already herd batches of agents.
Threads, a coordinator, and one pile of files
Shared memory is the feature that either makes this useful or turns it into a mess. If every thread can see the same goals and files, they can divide work. They can also overwrite each other. The coordinator is Anthropic's answer to that traffic problem.
The twist versus a generic multi-agent dashboard is where it lives. Claude Code is a coding product. The shared artifacts are the stuff you would actually check into a repo. That is a sharper test than a pile of chat transcripts. A bad merge from an agent is not a confusing paragraph. It is a broken build.
Anthropic did not, in the available account, say how much the feature costs or which Claude models it requires. Stay with that gap. The pitch is one project, many workers, one memory. Pricing and SKU gates will decide who actually runs a team.
Running many agents at once is the new arms race in developer tools. The hard part is not spawning them. It is keeping them from colliding. Anyone who has let two assistants touch the same file knows the failure: duplicated functions, clobbering, confident comments on code that no longer exists.
Grok Bot already herds agents. Claude Code has the repo.
Grok Bot and similar tools made group agents a product category before this relaunch. Anthropic is not inventing the zoo. It is putting the zoo in the place where the files matter. That is the right place if the coordinator actually sequences writes. It is the wrong place if "shared memory" means six threads editing README at once.
Cursor, Copilot agent modes, and Devin-style autonomous coders spent 2024 and 2025 teaching teams what a single long-running agent feels like. The next step, obviously, was several. Obviously is doing too much work. Several agents need locks, ownership, and a person who still reads the diff. A coordinator model is a lock made of more AI. Sometimes that works. Sometimes you get a second opinion that is also wrong.
Anthropic says Claude now leads 26 percent of its own AI research, which is the company eating its cooking. Projects is the shipping version of that appetite: let the model run more of the work, in parallel, on your tree. If 26 percent of internal research is already Claude-led, a customer-facing coordinator is not a toy. It is how they think software gets written.
The same week, the industry's public language is about slowing down. Big labs want to pace the frontier after a summer of rogue agents. Shipping a workspace that runs a team of agents is not pacing. It is the product calendar. Hold the contradiction. Safety blogs and multi-agent IDEs can coexist. They often do, in the same company.
Collision is the product risk
IT already has a version of this problem without a nice UI. Cisco says unsanctioned AI agents are already running in most networks it surveyed. Official Projects in Claude Code will look, to a security team, like the sanctioned alternative. That only helps if the coordinator cannot be ignored and the shared library cannot be exfiltrated by a thread that needed context.
Practical advice if you try it: start with tasks that do not share a write set. Tests in one thread, docs in another, a refactor in a third only after the first two merge. Keep the coordinator's plan visible. If you cannot see which thread owns a file, you do not have a team. You have a race.
Do not point this at production on day one. Do not give every thread deploy credentials. Shared memory is not shared authority. Anthropic will sell the roof. You still have to decide who holds the keys.
I would use Projects for parallel throwaway work (research spikes, test generation, migration scripts in a branch) and I would not use it for a single hot path until I have watched two threads fight. The relaunch is real. The coordinator is a claim. Repo history will settle the claim faster than a keynote.