Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Benchmark Finds MCP Task Decomposition Raises Token Cost 7.4 Times

AI summary

In a scripted MCP benchmark, splitting one weekly-report workflow into three short tools raised cumulative prompt tokens from 569 to 4,194 and execution time from 960 ms to 1,282 ms. The main source of the token increase was repeated transmission of intermediate results across turns. The decomposed workflow could retry only the failed step, but enforcing order required explicit state; without it, the protocol allowed a step to run out of sequence and return incomplete output. A task primitive used 560 prompt tokens and addressed server lifecycle rather than context costs. Because the experiment covered one three-step workload with a local scripted LLM, its results may not generalize to other tasks.

In-depth insight

The benchmark explicitly compares token usage, latency, recovery granularity, ordering-state costs, and task-lifecycle handling, showing that MCP decomposition involves tradeoffs beyond modularity. The report recommends measuring the relevant workload's own failure rate, step duration, intermediate-result size, and timeout constraints rather than applying these results universally.

上一篇结尾我给了个建议:长任务如果等不完,不如拆成几个短工具,让模型自己串起来。

理由写得很顺:不需要任务原语、不需要持久化 store、中途失败只重跑最后一步、模型还能看到中间结果改主意。代价我也提了一句——"模型得多调几次工具,多花点 token"。

这句话写得太轻了。

这次我真的拆了一遍,把账单量出来:同一个任务,拆成三个工具之后,累计 prompt 从 569 tokens 涨到 4194,7.4 倍。  而省下的恢复时间,只有 300 毫秒。

这篇文章就是把这三笔账摊开——token、时间、恢复代价——然后回答一个问题:什么时候拆值,什么时候拆是自残。

先说实验怎么做的

同一个「生成周报」的任务,三种形态,同一份数据(本周 5 篇笔记,7 个标签):

#MCP

Original · 11 min readRead the original
Group chat

No comments yet. Start the conversation.