大模型Agent

Agent

Agents are programs where LLM outputs control the workflow, they are useful when you need an LLM to determine the workflow of an app.

1. requirements

• 编译型 (Dify) 固定工作流 or 解释型 (Manus) 自主规划决策

• Router: LLM output determines an if/else switch

• Tool Caller: LLM output determines function execution

• Multistep Agent: LLM output controls iteration and program continuation

• Multi-agent: One agentic workflow can start another agentic workflow

• 记忆：长期记忆 or 短期记忆

2. ML task & Pipeline

• LLM (具备function call 能力)

• Tools: Plugins, Function Call, Code Interpreter

• Planning: CoT, ToT, ReAct

• Memory

• Self-Reflection / Self-Correction