The ReAct Pattern

ReAct (Reasoning + Acting) is a prompting strategy where the model explicitly generates:

1. Thought: Reasoning about the current state.
2. Action: The tool call to make.
3. Observation: The result of the tool call.

This loop continues until the agent decides it has enough information to answer.

Sruja Model

We can model this flow using a scenario or story in Sruja to visualize the sequence.

import { * } from 'sruja.ai/stdlib'


component Agent
component Tool
component User

story ReActLoop "Answering a Question" {
User -> Agent "Ask: What is the weather in SF?"

// Step 1
Agent -> Agent "Thought: I need to check weather"
Agent -> Tool "Action: WeatherAPI(SF)"
Tool -> Agent "Observation: 15°C, Cloudy"

// Step 2
Agent -> Agent "Thought: I have the answer"
Agent -> User "Answer: It's 15°C and cloudy."
}

view index {
include *
}

This visualization helps stakeholders understand the latency and cost implications of the multiple steps involved in a single user request.