Anthropic Just Dropped Agent Swarms

Key Takeaways

Agent teams = shared task list + inbox. Two primitives combine to form collaborative multi-agent sessions. The task list (with dependency blocking) handles work distribution; the inbox handles real-time communication between teammates.
Teammates persist until explicitly shut down. Unlike short-lived sub-agents spawned via the Task tool, teammates stay alive across tasks. You can follow up with individual teammates, add new tasks, or spawn additional teammates mid-session.
Inter-agent messaging solves the mergeability problem. Previously, parallel sub-agents worked on differing assumptions with no way to sync. The inbox lets a front-end agent and back-end agent share discoveries as they work, reducing merge conflicts.
Devil's advocate pattern. Spawn a critique agent alongside implementation agents. Real-time feedback during work produces better results than post-hoc review.
Opus 4.6 improvements. Better sub-agent orchestration (the model recognizes when to delegate), improved long-context performance, and adjustable effort levels (low/medium/high) to control cost and reasoning depth.
Auto-memory. Claude Code now persists learnings to a .claude memory directory per project. Useful in theory, but risks filling up with noise unless pruned regularly.

Architecture

The team system follows a leader-worker pattern with bidirectional messaging:

graph TD
    Leader[Team Lead] -->|creates| TL[Shared Task List]
    Leader -->|spawns| T1[Teammate A]
    Leader -->|spawns| T2[Teammate B]
    Leader -->|spawns| T3[Teammate C]
    T1 -->|claims task| TL
    T2 -->|claims task| TL
    T3 -->|claims task| TL
    T1 <-->|inbox messages| Leader
    T2 <-->|inbox messages| Leader
    T1 <-->|inbox messages| T2
    TL -->|blocks/unblocks| T3

Notable Quotes

"By combining these two pretty basic ideas—a shared task list with blockers and a direct message system between sub-agents—we now unlock more efficient parallel collaboration and can solve more complex problems."

"Someone on the Anthropic team made a C compiler using agent teams... after nearly 2,000 Claude Code sessions and $20,000 in API costs, agent teams produced a 100,000-line C compiler."

Connections

claude-codes-new-task-system-explained - Same author's previous video covering the task system that agent teams build upon—the shared task list with dependency blocking
fastrender-building-a-browser-with-thousands-of-ai-agents - Demonstrates the extreme end of multi-agent scaling: 2,000 concurrent agents producing a million lines of code
agentic-design-patterns - Covers the theoretical multi-agent collaboration patterns that Claude Code's team system implements in practice