Function: react()

@providerprotocol/agents

---

@providerprotocol/agents / execution / react

Function: react()

react(options): ExecutionStrategy

Defined in: src/execution/react.ts:78

Creates a ReAct (Reason-Act-Observe) execution strategy for agent execution.

The ReAct strategy implements the Reason-Act-Observe pattern from the paper “ReAct: Synergizing Reasoning and Acting in Language Models”. It adds an explicit reasoning phase before each action, which improves the model’s decision-making by encouraging step-by-step thinking.

The execution cycle is:

1. Reason: LLM outputs reasoning about what to do next
2. Act: LLM selects and executes tool(s) based on reasoning
3. Observe: Tool results are formatted as observations
4. Repeat: Continue until stop condition, no more actions, or maxSteps

This strategy is ideal for complex tasks requiring careful deliberation and multi-step reasoning.

Parameters

options

ReactOptions = {}

Configuration options for the ReAct strategy

Returns

ExecutionStrategy

An ExecutionStrategy instance implementing the ReAct pattern

Example

import { createAgent, react } from '@providerprotocol/agents';

// Basic ReAct agent
const agent = createAgent({
  llm: myLLM,
  tools: [searchTool, calculatorTool],
  strategy: react(),
});

// With custom reasoning prompt and step limit
const customAgent = createAgent({
  llm: myLLM,
  tools: [searchTool],
  strategy: react({
    maxSteps: 10,
    reasoningPrompt: 'Analyze the situation and determine the best action.',
  }),
});

// With lifecycle hooks for observability
const observableAgent = createAgent({
  llm: myLLM,
  tools: [searchTool],
  strategy: react({ maxSteps: 5 }),
  hooks: {
    onReason: (step, reasoning) => console.log(`Step ${step} reasoning:`, reasoning),
    onAct: (step, actions) => console.log(`Step ${step} actions:`, actions),
  },
});

See

• https://arxiv.org/abs/2210.03629 - ReAct: Synergizing Reasoning and Acting
• loop for simpler tool-loop execution
• plan for structured multi-step execution