Dispatching Parallel Agents

WHEN: The active agent is about to spawn more than one subagent in the same step. The temptation is to parallelize everything for speed. The reality is that parallel writes, parallel state mutation, and parallel coupled steps create races that single-threaded execution would have caught.

When to Use This Skill

Load this skill when:

• The agent is about to launch more than one subagent in the same step
• A task fans out into many independent reads (search, doc lookup, file inspection)
• A long analysis can be split into independent slices that each return a structured summary
• A research phase needs N viewpoints on the same artifact (council pattern)
• Skip when the work has any cross-subagent dependency, writes the same files, or mutates shared state

Prerequisites

• Knowledge of whether the work is read-only or write-bearing
• A bounded concurrency primitive available (e.g. tests/bounded-parallel-behavior.ps1, language-native parallel runner, or Start-ThreadJob with a throttle)
• A structured return contract for each subagent (JSON / Markdown section with a stable schema) -- free-form prose is hard to merge