你的 AI 代理在字段名称上浪费了 90% 的令牌

📄 中文摘要

在对系统中所有 31 个技能进行令牌审计后发现,AI 代理的令牌消耗主要集中在技能上,而非 API 响应。主代理 Francis 加载了 11 个技能,其中 'always:true' 技能在每条消息中都会注入,消耗了约 17,500 个令牌。包括按需技能在内,总令牌消耗约为 20,600 个。全局范围内,所有 31 个技能的 'always:true' 技能总共消耗了约 27,500 个令牌。这一发现表明,优化技能的令牌使用是提升效率的关键。

📄 English Summary

7º. Your AI agent is wasting 90% of its tokens on field names

A token audit conducted across all 31 skills in the system revealed that the token consumption of the AI agent was primarily due to the skills rather than the API responses. The main agent, Francis, loaded 11 skills, with 'always:true' skills injected into every message, consuming approximately 17,500 tokens. Including on-demand skills, the total token consumption reached around 20,600 tokens. Globally, the 12 'always:true' skills across all 31 skills consumed about 27,500 tokens. This finding indicates that optimizing token usage in skills is crucial for improving efficiency.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等