🎯
Coordinator Mode — Multi-Agent Orchestration
One Claude orchestrating a team of parallel worker agents
P1Core Pattern
Summary
Spawns multiple parallel worker Claude agents via a mailbox system. Workers communicate via <task-notification> XML messages. Shared scratchpad directory for cross-worker durable knowledge sharing. Explicit ban on lazy delegation: 'Do NOT say based on your findings — read the actual findings.'
Architecture Diagram
Technical Details
Three subagent execution models confirmed:
• Fork — independent tasks, no shared state
• Teammate — collaborative tasks, shared context (AsyncLocalStorage for isolation)
• Worktree — code branch-level isolation
Agent Teams/Swarm capabilities (tengu_amber_flint flag) include: in-process teammates, process-based teammates using tmux/iTerm2 panes, team memory synchronization, and color assignments for visual distinction.
Prompt instructs: "Parallelism is your superpower. Workers are async. Launch independent workers concurrently — don't serialize work that can run simultaneously."
Implementation Pattern
TypeScript (conceptual)
// Three subagent execution models
type SubagentMode = 'fork' | 'teammate' | 'worktree';
interface WorkerConfig {
mode: SubagentMode;
task: string;
// Fork: independent, no shared state
// Teammate: shared context via AsyncLocalStorage
// Worktree: git branch-level isolation
}
// Mailbox communication pattern
interface TaskNotification {
from: string;
to: string;
type: 'assign' | 'complete' | 'error';
payload: unknown;
}
async function coordinate(tasks: Task[]) {
// "Parallelism is your superpower"
const independent = findIndependentTasks(tasks);
const workers = independent.map(t => spawnWorker(t));
await Promise.all(workers); // Don't serialize!
}Architecture Insight
The three-tier model (Fork/Teammate/Worktree) maps to increasing levels of shared state: none, in-memory, and filesystem. This is a practical decomposition of the isolation-vs-collaboration tradeoff.
Official / Public Basis
Subagent documentation partially public (docs.anthropic.com/en/docs/claude-code/sub-agents). Coordinator Mode, mailbox system, and Agent Teams found in source.
Governance Concerns
Explicit ban on lazy delegation ('Do NOT say based on your findings'). Coordinator must read actual worker output. Shared scratchpad requires access control. Each worker needs clear accountability.
LightHope Ecosystem Mapping
LightHope AgentFactory — m8001/m8002/m8003/m8004 four-corner governance as agent roles, large project parallel execution, content + audit + deploy multi-agent pipeline