Parallel Research

A research agent using fork (all mode) with spawn-wrapped LLM calls to investigate a topic from multiple perspectives in parallel.

Overview

A research agent that investigates a topic from three perspectives simultaneously -- historical, technical, and societal. Demonstrates how fork fans out into parallel spawn-wrapped LLM calls and merges the results.

Primitives used: fork (all mode) + spawn + step.llm

How It Works

Input arrives as a research topic string
fork (mode: 'all') creates three parallel paths
Each path is a spawn-wrapped step.llm with a perspective-specific system prompt
All three complete, then merge combines results into a multi-section summary

Code

import { fork } from '@noetic/core';
import { spawn } from '@noetic/core';
import { step } from '@noetic/core';

const PERSPECTIVES = [
  {
    id: 'historical',
    label: 'Historical Context',
    instructions: 'You are a historian. Analyze the topic from a historical perspective.',
  },
  {
    id: 'technical',
    label: 'Technical Analysis',
    instructions: 'You are a technical expert. Analyze the topic from a technical perspective.',
  },
  {
    id: 'societal',
    label: 'Societal Impact',
    instructions: 'You are a social scientist. Analyze the topic from a societal perspective.',
  },
] as const;

export function buildParallelResearchAgent() {
  return fork<string, string>({
    id: 'parallel-research',
    mode: 'all',
    paths: () =>
      PERSPECTIVES.map((perspective) =>
        spawn<string, string>({
          id: `research-${perspective.id}`,
          child: step.llm<string, string>({
            id: `llm-${perspective.id}`,
            model: 'gpt-4o',
            instructions: perspective.instructions,
          }),
        }),
      ),
    merge: (results) => {
      const sections = PERSPECTIVES.map(
        (perspective, i) => `## ${perspective.label}\n\n${results[i]}`,
      );
      return `# Research Summary\n\n${sections.join('\n\n')}`;
    },
  });
}

Key Concepts

Fork All Mode

fork with mode: 'all' runs every path to completion and collects all results. The merge function then combines them. This is the right choice when you need every perspective -- unlike 'race' (first wins) or 'settle' (tolerates failures).

Spawn for Isolation

Each LLM call is wrapped in spawn, giving it its own context boundary. This means each perspective agent has independent token tracking, memory, and item logs. If one perspective fails, it doesn't corrupt the others.

Data-Driven Paths

The paths function generates steps from the PERSPECTIVES array. Adding a fourth perspective (e.g., economic) requires only adding an entry to the array -- no structural changes to the agent.