dispatching-parallel-agents

Identify independent problem domains, then dispatch one agent per domain with a focused scope, clear goal, constraints, and expected output. Each agent works concurrently without inheriting your session context, and you review and integrate their summaries afterward.

Use parallel agents when failures are independent—different test files, subsystems, or root causes with no shared state. Avoid parallel dispatch when failures are related, when you need full system context, or when agents would edit the same files.

Use file-based handoffs instead of pasting task text. Generate a task brief file per agent, pass only the file path, and have each agent write its report to a uniquely named file you read after completion.

Use a shared board file where each worker writes a heartbeat block with progress, next action, blockers, and file claims. A single coordination-only agent reads all heartbeats, renders the canonical board, detects file-claim collisions, and escalates true blockers.

A good prompt is focused on one problem domain, self-contained with all needed context, and specific about expected output. Include exact test names and error messages, state explicit constraints like not changing production code, and request a summary of root cause and fixes.