subagent-best-practices

Optimizing subagent delegation requires enforcing a two-tier architecture where Claude acts as the orchestrator and cost-effective providers like Kimi K2.5, GLM, or MiniMax function as workers. This structure prevents inefficient token usage and reduces operational overhead.

Avoiding Claude models like Haiku, Sonnet, and Opus as subagents prevents high execution costs and enforces the recommended two-tier architecture. Using allowed alternative providers such as Kimi K2.5 or GLM ensures efficient parallel task execution without sacrificing workflow quality.

Selecting a provider for AI agent delegation involves matching the task type to recommended models like Kimi K2.5, GLM, or MiniMax. This guidance ensures optimal provider selection while maintaining the orchestrator and worker roles within the two-tier architecture.

Implementing background task suspension and parallel execution involves using a fire-and-retrieve pattern for subagent calls. This approach allows the orchestrator to dispatch multiple tasks simultaneously, saving tokens by retrieving results only when background processing completes.

A standardized prompt template is provided to structure subagent calls effectively. Using this template ensures clear communication between the orchestrator and worker agents, maintaining consistency across parallel execution and background task delegation workflows.