agent-expenditure

This Skill addresses the issues of uncontrolled token consumption and excessive coordination overhead when scheduling multiple AI agents in parallel. In real-world development, developers often dispatch multiple agents simultaneously to perform tasks but lack fine-grained tracking of each agent's token usage, leading to soaring costs and redundant outputs. Through post-hoc analysis, this Skill quantifies the token expenditure of each agent and flags wasteful behavior based on Brooks's law, helping teams optimize the number of agents.

Usage is straightforward: after parallel agent execution completes, simply trigger the Skill via a natural language instruction for post-analysis. It automatically checks whether each agent's output is unique, whether token expenditure is proportional to value, and generates a report including coordination overhead percentage and agent count recommendations. You can then evaluate each item in the report's "Post-Dispatch Review Checklist" to decide whether to reduce the number of agents in future dispatches.

It is ideal for teams or individuals managing multiple AI agents, especially those already experiencing skyrocketing token costs, agent redundancy, or inefficient coordination. Project managers can use it to quantify scheduling efficiency, agent developers can adjust parallel strategies accordingly, and operations engineers can monitor resource consumption. For single-agent workflows or real-time token budget control scenarios, this Skill is not needed.

It is recommended to run this Skill uniformly after each large-scale parallel dispatch for retrospective analysis, combined with the "plan-before-large-dispatch" rule to pre-plan agent counts. Note that it is primarily for post-hoc analysis, does not support real-time interception or budget control, and relies on the Claude Code plugin environment. Additionally, when the number of agents exceeds nine, coordination overhead may surpass 30%, in which case reducing agents should be prioritized over adding more.