Operating rules for running multi-subagent fleets in Claude Code. Covers model tiering by task, per-agent reasoning effort, concurrency discipline, verification doctrine the fleet inherits, a two-seat check for decisions you can't cheaply redo, and time-boxed "work until X" mandates. Use when fanning out subagents, running the Workflow tool (agent(), parallel(), pipeline()), planning a multi-agent build, choosing a model or effort level for an agent, spawning an orchestrator, making a call that can't be cheaply redone, or given deadline-shaped autonomous work.
