Descriptions:
Claude Code’s Agent Teams feature introduces two core primitives for multi-agent workflows: a shared persistent task list with dependency blocking, and a direct inbox messaging system between subagents. Ray Amjad explains how combining these allows subagents to communicate assumptions, share findings mid-task, and unblock each other — solving the merge conflicts and silent assumption divergence that plagued earlier multi-agent setups where agents had no way to exchange information while working.
The video walks through enabling Agent Teams with the required environment variable, understanding the two display modes (concurrent and split-pane), and reading the `.claude` folder structure that stores task JSON files and per-agent inboxes. A practical example shows a team checking feature parity between macOS and Windows versions of Amjad’s HyperWhisper app, with a planner and implementer coordinating through the inbox — each team member claiming tasks, marking completions, and routing findings back to the team lead.
A more advanced pattern demonstrates running a “devil’s advocate” subagent in parallel with a designer agent using tmux inside Ghostty or iTerm2, providing real-time critique rather than end-of-task review. Amjad shows how to send direct messages to individual team members while they work, how to resume a team session after exiting Claude Code, and how to request a clean shutdown. The result is a coordination model that more closely mirrors real engineering teams — with handoffs, blockers, and parallel execution — inside a single Claude Code environment.
📺 Source: Ray Amjad · Published February 06, 2026
🏷️ Format: Tutorial Demo
