Enter your search text in the box above

Select a result to preview

LangChain ReAct Agent Loop (Legacy)

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

LangChain ReAct Agent Loop (Legacy)

Deprecated at 0.1.0, removed at 1.0. Defined in langchain_classic/agents/react/base.py.

Implements the ReAct paper (arXiv 2210.03629): interleave Reasoning (Thought) and Acting (tool calls) in a single text stream.

Prompt template: the control protocol lives in text

The loop is driven by two template variables filled into a PromptTemplate:

Variable Role
{input} The question/task — filled once at invocation
{agent_scratchpad} Growing string of Thought:/Action:/Observation: turns 
SUFFIX = """\nQuestion: {input}
{agent_scratchpad}"""
WIKI_PROMPT = PromptTemplate.from_examples(EXAMPLES, SUFFIX, ["input", "agent_scratchpad"])

from_examples() prepends static few-shot demonstrations showing the Thought/Action/Observation/Finish pattern; the LLM learns the protocol from them.

Control signals — all text

Signal Form Who produces it
Tool call Action: ToolName[argument] e.g. Search[Colorado orogeny] LLM (plain text)
Exit Action: Finish[answer] LLM (plain text)
Stop generation stop token "\nObservation:" Framework (injected as stop sequence)
LLM prefix "Thought:" Framework (llm_prefix property)
Result injection "Observation: <result>" appended to scratchpad Framework

ReActOutputParser reads Action: / Action Input: lines from the LLM's text to determine what to execute. There is no structured object — the entire control protocol is a text convention enforced by the prompt format and the parser together.

Loop mechanics

The loop lives in the parent AgentExecutor class, not in ReActDocstoreAgent itself:

┌──────────────────────────────────────────────────────┐
│  Fill {input} + {agent_scratchpad} into template     │
│  LLM call with prefix "Thought:", stop "\nObs:"      │
│  ReActOutputParser reads Action: ToolName[arg]       │
│  Execute tool → observation string                   │
│  Append "Observation: <result>" to scratchpad        │
│  Repeat until LLM writes "Action: Finish[answer]"   │
└──────────────────────────────────────────────────────┘

Class hierarchy

ReActDocstoreAgent(Agent)
    ↑ used by
ReActChain(AgentExecutor)   ← thin wrapper that wires DocstoreExplorer

Tool constraints

Validated at construction; hard-coded to exactly {"Lookup", "Search"}. The DocstoreExplorer helper bridges these to an abstract Docstore:

Tool What it does
Search Retrieves a document; saves it as current document
Lookup Finds next occurrence of a term in the saved document (paginated)

DocstoreExplorer is stateful: it tracks the current document and lookup cursor across turns.

Comparison

See Source - LangChain Agent Loop for full legacy vs modern comparison table.

Modern replacement: LangChain LangGraph Agent Loop.