Dispatching Parallel Agents

Dispatching Parallel Agents

Overview

When you have multiple unrelated failures or investigations, handle them in parallel instead of forcing one agent to juggle everything.

Core principle: one agent per independent problem domain.

When to Use

• Multiple failures with different root causes
• Separate subsystems or files
• No shared write set
• No sequential dependency between the tasks

The Pattern

1. Identify the independent domains.
2. Give each agent a narrow, self-contained prompt.
3. Dispatch them in parallel.
4. Review the results and integrate carefully.

Codex CLI Example

ext spawn_agent(role=explorer, task="Investigate agent-tool-abort.test.ts failures") spawn_agent(role=explorer, task="Investigate batch completion behavior failures") spawn_agent(role=explorer, task="Investigate tool approval race conditions")

Common Mistakes

• One prompt that covers unrelated problems
• Missing context or constraints
• Parallel work on overlapping files
• No integration pass after the agents return