ReAct Pattern

ReAct (Reasoning + Acting) is the foundational agent loop. The LLM alternates between reasoning about the task, calling a tool, and observing the result. Introduced by Yao et al. (2022), it is now the default agent pattern in LangGraph.

The Loop

1. THINK  — LLM reasons about what to do next (internal chain-of-thought)
2. ACT    — LLM emits a tool call with arguments
3. OBSERVE — Tool executes, result is appended to message history
4. REPEAT — Loop back to THINK until LLM produces a final answer (no tool call)

LangGraph Implementation

LangGraph implements ReAct as a two-node graph with a conditional edge:

[LLM Node] → has tool calls? → YES → [Tool Node] → back to [LLM Node]
                              → NO  → END

from langchain.agents import create_agent

agent = create_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[search, calculate, format_output],
    system_prompt="You are a helpful assistant."
)

The create_agent() function (previously create_react_agent()) builds this graph, binds tools to the model, and handles the conditional routing.

Key Parameters

model — LLM identifier or instance
tools — list of tool functions
system_prompt — agent instructions
response_format — structured output schema (optional)
state_schema — custom state beyond messages (optional)
middleware — hooks for modifying model requests/responses
name — identifier for multi-agent systems

Practical Limits

ReAct works well with 1-12 tools in a single domain for tasks requiring <10 tool calls. Beyond these thresholds, consider AI Agent Architectures|multi-agent patterns:

>12 tools: LLM struggles to select correctly. Use skills or subagents.
>10 sequential tool calls: Context bloat. Use subagents for isolation.
Multiple domains: Single prompt can’t specialize. Use handoffs or subagents.
Complex planning: ReAct has no explicit planning step. Consider plan-then-execute patterns.

In kulify Projects

fajb-next: createReactAgent() with 12+ tools for article editing
master-bot: Simple agent loop with availability tools
Vorma (Furniq): Tools return __stateUpdate for 3D config