Claude Code Subagent Cost Optimization

According to Anthropic's data, multi-agent systems use about 15x the tokens compared to a single chat, but leveraging Prompt Caching can reduce costs to around 10%. The key to cost savings lies in whether all subagents share the same prefix.

Three delegation methods by cost and efficiency:

• subagent_type configuration: Provides an isolated environment using custom prompts and tools. However, since caching cannot be shared, full cost is incurred every time.
• Parent agent clone: Inherits the parent's prompt and tools as-is. Since child agents leverage the cache (10% pricing), running tasks in parallel can increase cost by about 1.5x while making things 5x faster.
• Keeping the main agent: For dependent tasks where step-by-step results are needed for the next step (e.g., refactoring), this is the cheapest and most effective option.

Recommendations by task type:

• Good: Tasks that allow for parallel research, such as reading multiple files in parallel, auditing patterns within a folder, or gathering information from various sources.
• Bad: Sequential tasks with high step-by-step dependency, such as bug fixes or module refactoring.