ReAct (Reasoning + Acting)
Definition
ReAct ist ein Paradigma, bei dem das Modell Reasoning abwechselt (was als Nächstes zu tun ist, warum) and acting (Tool-Aufrufe). The observation aus dem environment feeds back into the next Schlussfolgern step, forming a loop until the task is done.
Es ist the standard pattern for agents that use tools: each action is preceded by a thought, was reduziert blind or repetitive tool use. Often combined with chain-of-thought (Schlussfolgern inside the thought) and with RDD when specs guide Entscheidungs.
Funktionsweise
Prompt-Format ist Gedanke → Aktion → Beobachtung → Gedanke → … → Endgültige Antwort. Der Benutzer gibt eine Aufgabe; der Agent erzeugt a thought (Schlussfolgern about what to do), then an action (z. B. tool call). The environment/tools return an observation, which is appended to the context für den next thought. The loop continues until the agent outputs a final answer. The model decides when to call tools and when to conclude, was reduziert arbitrary or repetitive actions. The sequence diagram below summarizes this flow; frameworks like LangChain implement ReAct-style agents with tool registration and message handling.
Anwendungsfälle
ReAct fits agent workflows where each tool call should be preceded by a clear Schlussfolgern step.
- Agents that use tools (search, calculator, API) with explicit Schlussfolgern
- Reducing arbitrary or repetitive Tool-Aufrufe by interleaving thought
- Debuggable agent behavior via visible thought–action–observation traces
Externe Dokumentation
- ReAct: Synergizing Reasoning and Acting in LLMs (Yao et al.) — Original ReAct paper
- LangChain – ReAct agent — ReAct-style agents in LangChain