DecisionFlow Mini
把证据、路由、审核和审计接成一条可追溯的决策链。
AI Engineer / Technical notebook
记录 Agent systems、workflow、evaluation 与生产边界。这里不追逐“更像 Agent”的表演,而是拆解一个系统为什么可靠、哪里会失败,以及工程师应该为哪些判断负责。
把证据、路由、审核和审计接成一条可追溯的决策链。
让 Agent 的运行状态、证据、恢复和治理成为一等对象。
一个把浏览器自动化、岗位判断、人工接管和失败记录放在一起的开源工程实践。
A point of view
我把 LLM 看作不确定的决策部件,把 runtime、权限、状态、评估和人工接管看作让它能够进入现实的工程边界。
读我的判断 ↗The value of agent engineering is not giving every step to the model
A first-person argument about where agent engineers create durable value when general-purpose agents can already use tools and operate software.
For agents, context is often closer to behavior than workflow
Why I treat context, memory, identity, and permissions as runtime foundations instead of assuming that orchestration alone explains agent behavior.
Workflow vs agent loop: do not hide control ownership behind one word
A practical distinction between deterministic workflows, graph runtimes, and LLM-led control policies, based on comparing real agent architectures.