Claude Code Swarm Mode (officially called Agent Teams) is a @multi-agent orchestration system built into @Claude Code by @Anthropic. Initially feature-flagged and hidden from users, the system was discovered by the developer community through code analysis before being officially documented. Agent Teams transform Claude Code from a single AI assistant into a coordinated multi-agent system.

How It Works

The primary Claude Code agent can spawn specialized teammate agents and delegate work through natural language instructions. Each teammate runs in its own context window with:

Independent tool access — each agent has its own permissions and available tools
Shared task list — coordination primitive for tracking what's assigned and what's complete
Peer-to-peer messaging — agents can communicate directly with each other, not just with the orchestrator
File locking — prevents merge conflicts when multiple agents edit the same codebase This enables parallel task execution: one agent refactors the backend while another updates tests while a third writes documentation — all coordinated through the shared task list.

How It Compares to Sub-Agents

Claude Code already supported sub-agents before Agent Teams. The difference:

Sub-agents run in isolation. They receive a task, complete it, and return a result to the parent. No peer communication, no shared state
Agent Teams add coordination primitives. Shared task lists, peer messaging, and file locks make them suitable for tasks where agents need to be aware of each other's work For most tasks, sub-agents are sufficient and simpler. Agent Teams are valuable when multiple agents need to modify the same codebase simultaneously — large refactors, feature branches with parallel workstreams, or comprehensive test suite additions.

Connection to BrianBot

@BrianBot's 57-agent system predates Agent Teams, using @OpenClaw's session spawn system for orchestration. Agent Teams represents Anthropic building native multi-agent coordination into Claude Code itself — validating the pattern that practitioners like Brian have been running in production. The coordination primitives (shared task lists, peer messaging, file locks) mirror what BrianBot built with custom infrastructure. See: @The BrianBot Architecture. As Agent Teams matures, it may complement or partially replace OpenClaw's orchestration for @Claude Code-native workflows, while OpenClaw continues handling multi-channel agent coordination (Slack, WhatsApp) that Agent Teams doesn't address. When Agent Teams leaked, my first reaction was recognition — Anthropic built the same coordination primitives I'd been running in @BrianBot for months. Shared task lists, peer messaging, file locks. The patterns are the same because the problems are the same. At scale, @multi-agent orchestration is a coordination problem, not an AI problem. Anthropic discovered this the same way I did: by building it.

Contexts

#agentic-augmentation
#anthropic
#claude-code