13. Multi-Agent Value Is Not Multiple Models Running
Concurrency is scheduling; the engineering value of a child Agent comes from an explicit task boundary
系列专题
Agent Runtime Lab: From Packet Capture to Replay
16 篇文章，按顺序阅读或从当前问题进入。
14. Recording a Replayable Agent Trace
Replay reconstructs decisions and external actions in causal order; it is not a second rendering of the final answer
05. What Happens After the Model Calls a Tool
A model tool call is an execution request, not evidence that the action has happened
03. Where Prompt Injection Enters an Agent
Text entering context is only the first step; impact depends on source handling, permission checks, and side-effect boundaries
11. Which Layer Should Handle an Agent Error?
A 403, an invalid argument, a denied operation, and a failed process all stop a task but require different handling
08. Why an Agent Needs a State Machine
Conversation records what the model saw; state records what the runtime has actually completed