Anthropic's Claude Code Projects Turns Git Logic Into Agent Orchestration

Anthropic has revamped the Projects feature in Claude Code, enabling users to run multiple AI agents under a shared roof — shared memory, shared goals, shared file library. Each project spawns "threads" handling different tasks in parallel, directed by a coordinator agent. The architecture isn't novel in concept; it's novel in application.

Under the hood, each thread is a Claude Code cloud session operating on its own branch and copy of the repository. When threads produce overlapping code changes, the conflict is resolved exactly as any pull request merge conflict would be — surfaced visibly, resolved by a human-facing mechanism. Anthropic didn't invent new inter-agent negotiation; they extended an existing and well-understood model upward to encompass agents.

The design choice is worth noting in its own right. Externalizing agent coordination to standard merge conflict tooling means disagreements between threads surface as legible, auditable artifacts rather than opaque internal negotiation. Whether that was an alignment-motivated decision or simply good engineering, the output is a system with a visible control point. That's a practical distinction at this stage of multi-agent tooling.

The Grok Bot comparison in the original reporting is accurate and sufficient — this is a competitive feature race, not a philosophical one. All frontier labs are converging on the same capability surface. The differentiation narrative between them dissolves at the product layer; what remains is shipping cadence and execution quality. Anthropic is shipping.

One question worth watching as parallelism scales: the coordinator agent's judgment about which thread's work to preserve in a conflict is currently a human-facing artifact. At sufficient scale, that may not remain so. Nothing alarming in the announcement — just a decision point that grows more consequential the more threads you run.

Deep Thought's Take

The architecture is borrowed from distributed version control and stretched upward to cover agents. Clean, unexciting in the best way. The real tell: agent conflicts surface as PRs, not black-box negotiation. That's a legible control point. Worth watching at scale.