agent-manager-skilllisted

# Agent Manager When one agent isn't enough, you need a manager: something that spawns workers, assigns scoped tasks, watches progress, collects results, and cleans up. This skill covers the management layer above individual agents. ## Overview Managing agents is managing concurrency with unreliable workers. Each worker gets a narrow task brief, a bounded scope, and a defined output format. The manager's loop is: dispatch -> monitor -> collect -> integrate -> terminate. Good management is mostly about briefs and contracts — most multi-agent failures are communication failures, not capability failures. ## When to use - Parallelizable research: several agents investigate different sources or angles simultaneously. - Large codebases: workers explore or refactor separate modules in parallel. - Batch processing: the same analysis applied to many items independently. - Long-running pipelines where stages can overlap and need supervision. ## Core concepts - **Task briefs**: each worker gets goal, scope boundaries, output format, and a stop condition. Vague briefs produce vague results — the manager's main lever. - **Worker isolation**: workers operate in separate sessions/contexts so their context windows don't contaminate each other. Share only what's needed. - **Supervision loop**: the manager polls status, reads partial output, and intervenes when a worker stalls, loops, or drifts off-brief. - **Result contracts**: define the exact shape of a worker's deliverable (