agent-team-orchestrationlisted

# Agent team orchestration The `team-lead` subagent (`.claude/agents/team-lead.md`) coordinates multi-phase work. **It runs in one of two modes depending on what your Claude Code build supports.** Pick the right mode for your situation, and frame your expectations accordingly. ## Mode A — Hub-and-spoke (standard Claude Code, available everywhere today) This is the mode you have unless you've explicitly opted into experimental features. **How it works**: the main agent calls `Agent({ subagent_type: "team-lead", ... })` once. The team-lead runs as a single deep sub-agent invocation. Inside that one run, it: 1. Decomposes the work into phases 2. Recursively spawns `feature-builder` for each phase (uses the `Agent` tool from within its own run) 3. Spawns `code-reviewer` after each phase 4. Runs verification (`tsc`, `npm test`, lint) between phases via Bash 5. Returns ONE final summary to the main agent **What you get**: - Recursive delegation that the user doesn't have to coordinate manually - Phase-boundary verification inside one deep call - A clean final summary the main agent reports up to the user - The main agent's context stays clean (the team-lead absorbs the implementation noise in its own context window) **What you don't get**: - Multi-turn persistence — the team-lead returns ONCE; it doesn't span multiple user prompts - The ability to pause for user input mid-stream and resume - Long-running sessions that survive across conversation turns **This is enough for**: