Control flow decides what happens next, what evidence changes the plan, and when the task ends. It can mix fixed code paths with model-selected actions. ReAct provides primary evidence for alternating reasoning and environmental actions; it does not establish that one visible reasoning format is required for every agent. react
Durable execution preserves enough state to resume a task after interruption. LangGraph distinguishes thread checkpoints from cross-thread stores and notes that in-memory checkpointers do not survive process restart. A saved conversation is only one part of recovery. langgraph-persistence
An agent system turns a task into actions and checks their effects. This knowledge base uses agent for a system in which a model can choose subsequent actions from observations, and Agent Harness for the surrounding software that runs that process. These are editorial working definitions: industry usage overlaps, and a framework or hosted runtime may supply only part of the harness.
Multi-agent coordination allocates work among model-driven workers and reconciles their outputs. Delegation is a usable engineering pattern; reliable autonomous decomposition and coordination remain sensitive to task coupling. The core label here applies to the design problem and trade-offs, not to a claim that multi-agent systems are generally preferable.