kbwen / dispatching-parallel-agents

Dispatching Parallel Agents

Overview

Parallel agents are suitable for handling splittable tasks with low coupling. The goal is to shorten total lead time, not to increase coordination overhead.

When to Use

• Tasks can be decomposed into independent sub-problems (e.g., documentation, testing, separate modules).
• Dependencies between sub-tasks are low; clear interfaces can be defined.
• Deadlines are tight, and lead time needs compression.

When NOT to Use

• Problems are highly coupled and require frequent synchronization.
• Requirements are still unclear; the cost of decomposition outweighs the benefits.
• Unable to define unified acceptance criteria.

Four-Step Pattern

1. Decompose: Split into mutually exclusive, independently verifiable sub-tasks.
2. Dispatch: Specify each agent's scope, inputs/outputs, constraints, and deadline.
3. Sync: Perform brief synchronizations at fixed checkpoints to handle conflicts and dependencies.
4. Merge: Integrate outputs using a unified set of acceptance criteria and run regression tests.

Best Practices

• Every sub-task MUST have a clear Definition of Done.
• Collaborate primarily via interface contracts to avoid modifying each other's blocks.
• Perform a global check before final integration.

Common Mistakes

• Forcefully splitting non-splittable tasks, leading to rework.
• Lacking sync checkpoints, resulting in integration explosion at the end.
• Sub-tasks passing individually but failing global acceptance.