orchestrating-swarms_skill

Overview

This skill helps you orchestrate multi-agent swarms using Claude Code's TeammateTool and Task system. It codifies primitives and patterns for spawning teammates, managing shared task lists, messaging via inboxes, and building divide-and-conquer pipelines. Use it to coordinate parallel reviews, dependency-driven pipelines, and persistent worker teams.

How this skill works

The skill maps swarm primitives (Leader, Teammate, Team, Task, Inbox, Message, Backend) to concrete operations you run via Task() and Teammate() calls. You can spawn short-lived subagents for focused work or persistent teammates that join a team and communicate through inbox JSON files. Tasks live under a team task directory, can have dependencies, and teammates claim and complete tasks while the leader coordinates via TeammateTool operations (spawnTeam, write, broadcast, approveJoin, requestJoin, shutdown).

When to use it

• Parallel code reviews across multiple specialized reviewer agents
• Create pipeline workflows with task dependencies and staged teammates
• Run background exploration or research agents that report findings to a leader
• Break large problems into independent subtasks for divide-and-conquer
• Manage long-lived worker agents that persist until explicit shutdown

Best practices

• Choose Task() without team_name for short, synchronous queries and searches; use teammates for persistent, background work
• Define clear task subjects, owners, and dependencies to enable automatic unblocking and pipeline flow
• Use explicit write() to message a single teammate; avoid broadcast() unless necessary to reduce message cost
• Pick appropriate built-in or plugin subagent types (Explore for read-only, general-purpose for multi-step tasks, specialized reviewers for audits)
• Prefer run_in_background for persistent teammates and ensure leader handles join/approve and graceful shutdowns

Example use cases

• Spawn a team of security-sentinel and performance-oracle agents to run parallel PR audits and aggregate findings
• Create a pipeline: explorer locates files -> implementer teammate applies changes -> reviewer teammates run targeted reviews, with tasks chaining via dependencies
• Run background git-history-analyzer agents to produce repository evolution reports while other teammates implement fixes
• Use research agents to gather current best practices, then feed findings into design and plan agents for architecture decisions
• Spin up temporary subagents to run quick searches or scripts without adding them to a team