Descriptions:
Claude Code version 2.1.34 shipped an experimental Agent Teams feature that goes well beyond the existing sub-agent capability — and AI Jason reverse-engineered the actual LLM call logs to explain precisely how it works. Unlike sub-agents that report only a summary back to the orchestrator, agent teams maintain a shared task list, can communicate peer-to-peer, and coordinate dependencies before picking up work.
The video traces the full lifecycle: the main agent calls `team_create` to initialize a team config under `.claude/teams/`, then `task_create` to populate a JSON-based task list with status, owner, and blocking relationships. From there, the familiar `task` tool spins up individual Claude Code sessions — but each session inherits team context and gains access to `task_update`, `send_message` (supporting direct and broadcast modes), and a `shutdown_response` handshake that lets the team lead cleanly terminate sessions via `team_delete`. Inter-agent messages are implemented by injecting new user messages into each agent’s conversation history through a shared inbox folder with read-status tracking.
Practically, AI Jason walks through enabling the feature with a single environment variable (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` in settings.json), and recommends tmux or iTerm2’s Python API with `claude –teammate-mode` for a split-screen view of all running sessions. The video is the clearest available reference for developers trying to understand when agent teams make sense versus sub-agents and how to structure tasks to take advantage of parallel coordination.
📺 Source: AI Jason · Published February 07, 2026
🏷️ Format: Deep Dive
