Agent Workflows: ReAct, Plan-and-Execute and Reflection

ki
llm
agenten
English

Agent Workflows: ReAct, Plan-and-Execute and Reflection

An agent combines an LLM with tools to solve tasks autonomously. The workflow — the loop of thinking and acting — decides reliability.

ReAct

Reason + Act: the model thinks aloud (Thought), calls a tool (Action), observes the result (Observation) and repeats until the task is done. This pattern is at the core of most agent frameworks.

Plan-and-Execute

First create a plan (step list), then execute the steps — with re-planning on deviations. Good for long, multi-step tasks, saves tokens compared to ReAct.

Reflection

The model evaluates its own output, finds errors and improves. Especially useful for code, text and decisions.

Tools

Web search and extraction.
Code execution (sandbox).
File access, APIs, databases.