Managing complex workflows with AI agents often leads to confusion, especially when one agent forgets previous requests amidst a sea of logs. The concept of subagents addresses this by allowing users to create specialized agents that can handle focused tasks, keeping the main conversation cleaner and more organized. By defining subagents in a markdown file, you can spawn them for specific functions, like searching or verifying code, without cluttering the main context. Subagents offer several advantages, including the ability to work in parallel and maintain distinct contexts for different tasks. This is particularly useful when multiple types of work need attention simultaneously, like scanning modules for issues or gathering design options. However, subagents also come with risks. If poorly managed, they can generate excessive output, leading to further confusion. Establishing clear output contracts and responsibilities is essential to minimize overlaps and ensure that each agent operates effectively. Setting limitations on how many subagents run in parallel is crucial to avoid rate limits and unnecessary token expenditure. A good starting point is to limit concurrent subagents to three. When you define their roles clearly and enforce strict contracts on what they return, you can streamline the workflow and reduce the likelihood of errors. Subagents should be treated as proposals rather than final solutions, with a verifier in place to confirm their outputs. This structured approach helps maintain order in a system that can easily become chaotic.