ReAct Pattern
The ReAct pattern (Reason + Act) prompts the model to reason before each tool call. This improves reliability on multi-step tasks.
Flint ships a react() recipe that implements the pattern. Use it as a drop-in replacement for agent() when you need better reasoning on complex tasks.
ts
import { tool } from 'flint';
import { react } from 'flint/recipes';
import { budget } from 'flint/budget';
import { anthropicAdapter } from '@flint/adapter-anthropic';
import * as v from 'valibot';
const adapter = anthropicAdapter({ apiKey: process.env.ANTHROPIC_API_KEY! });
const searchWeb = tool({
name: 'search_web',
description: 'Search the web and return a summary of results',
input: v.object({ query: v.string() }),
handler: async ({ query }) => `Results for "${query}": [mocked results]`,
});
const readPage = tool({
name: 'read_page',
description: 'Fetch and read the content of a webpage',
input: v.object({ url: v.string() }),
handler: async ({ url }) => `Content of ${url}: [mocked content]`,
});
const out = await react({
adapter,
model: 'claude-opus-4-7',
messages: [
{
role: 'user',
content: 'Find the current TypeScript version and summarize the major new features.',
},
],
tools: [searchWeb, readPage],
budget: budget({ maxSteps: 12, maxDollars: 1.00 }),
});
if (out.ok) {
console.log(out.value.message.content);
}How it works
react() injects a system prompt that instructs the model to emit a Thought: prefix before each action, then an Action: call. This structured reasoning is stripped from the final output before returning. The model's reasoning trace is available in out.value.steps.