Posts tagged with "MCP"

Wrapping device operations as tools an agent can call: the three layers, why a tool description is an inference rule rather than documentation, and the two shells a deeplink has to survive.

把裝置操作包成 agent 能呼叫的工具，中間踩到的兩層 shell、被悄悄蓋掉的 timeout，以及為什麼 tool description 寫完還不夠。

A few weeks of hands-on notes on agent runtime and agent identity - without a defined enforcement boundary, even a good agent identity design is just more metadata

這是我過去幾週針對 agent runtime & agent identity 的一些實作心得跟看法 - 沒有 enforcement boundary 定義，再好的 agent identity 設計也只是新增更多的 metadata

當你的 AOSP codebase 大到人腦裝不下，Claude Code + 自建 MCP server 能幫到什麼程度？這是我從「能動就好」到 16 個工具的完整過程，包括哪些假設被打臉。