Enter your search text in the box above

Select a result to preview

LangChain ReAct Prompt Templates

#langchain #react #prompt-template #agent-loop #legacy

Source: LangChain ReAct Prompt Templates

Raw file: raw/docs/langchain-react-prompts.md
Origin: langchain-ai/langchain (master)

What these files are

The actual prompt templates used by ReActDocstoreAgent and ReActTextWorldAgent. Both are deprecated alongside their agents (0.1.0 / removed 1.0). They make concrete what the code in base.py only implies: that the loop's control protocol is entirely a text convention.

Template variables

Both prompts share the same two input variables:

Variable Role
{input} Task or question — filled once at invocation
{agent_scratchpad} Growing string of Thought:/Action:/Observation: turns — appended to each iteration 
# wiki_prompt.py
SUFFIX = """\nQuestion: {input}
{agent_scratchpad}"""
WIKI_PROMPT = PromptTemplate.from_examples(EXAMPLES, SUFFIX, ["input", "agent_scratchpad"])

# textworld_prompt.py
SUFFIX = """\n\nSetup: {input}
{agent_scratchpad}"""
TEXTWORLD_PROMPT = PromptTemplate.from_examples(EXAMPLES, SUFFIX, ["input", "agent_scratchpad"])

PromptTemplate.from_examples() concatenates static few-shot demonstrations (showing the full Thought/Action/Observation/Finish cycle) followed by the dynamic SUFFIX. The few-shot examples teach the LLM the text protocol.

Control signals — all plain text

Signal Text form Who produces it
Tool call Action: ToolName[argument] LLM (plain text)
Exit Action: Finish[answer] LLM (plain text)
Stop generation stop token "\nObservation:" Framework (injected)
LLM turn prefix "Thought:" Framework (llm_prefix property)
Result injection "Observation: <result>" Framework (appended to scratchpad)

Tool names appear literally in text:

ReActOutputParser reads Action: / Action Input: lines from the LLM's output via string matching to determine what to execute next.

The prompt-as-protocol insight

The entire agent loop control protocol is a shared text contract between:

  1. The prompt template ({agent_scratchpad} slot)
  2. The few-shot examples (demonstrate the format)
  3. The output parser (ReActOutputParser reads Action: lines)
  4. The framework loop (injects stop tokens, appends observations)

There is no structured object. The LLM doesn't "call" a tool — it writes text that looks like a tool call, and the parser interprets it. This means:

Contrast with modern tool-call loops

Modern loops (LangGraph, pi-mono, Hermes) replace this entirely: the LLM emits a structured tool_calls array in its JSON response. The framework reads AIMessage.tool_calls directly — no text parsing, no stop tokens, no scratchpad string management. Tool invocation becomes a typed object; routing becomes a function on that object. The text convention is gone.