Deep Research Agent

createDeepAgent packages planner + parallel sub-agents + compression into a single opinionated factory for long-horizon research tasks.

import { createDeepAgent } from 'personaforge/skills';

---

Quick start

const deep = createDeepAgent({
  generate: (prompt) => llm.generate(prompt),
  tools: [webSearchTool, wikipediaTool],
});

const result = await deep.run('What are the long-term economic effects of UBI?');
console.log(result.answer);
console.log(result.subQuestions);
console.log(result.rawSubAnswers);
console.log(result.steps.map((s) => `${s.phase}: ${s.detail}`));

---

Pipeline

1. Plan — the LLM decomposes the question into focused sub-questions.
2. Research — each sub-question runs in parallel. Optional tools (search, Wikipedia) are called first, and their results are injected into the research prompt.
3. Synthesize — the findings are concatenated and a final synthesis prompt produces a structured answer with inline citations.

---

Configuration

interface DeepAgentConfig {
  generate: (prompt: string) => Promise<string>;  // any LLM
  tools?: Array<{ name; description; execute }>;   // called per sub-question
  maxParallel?: number;      // default 5
  maxQuestions?: number;     // default 5
  subAnswerMaxChars?: number; // default 2000
}

---

Result shape

interface DeepResearchResult {
  answer: string;                                  // multi-paragraph synthesis
  steps: Array<{ phase; detail }>;                 // audit trail
  subQuestions: string[];
  rawSubAnswers: Array<{ question; answer }>;
}

---

Usage tips

• Narrow maxQuestions for simple queries to avoid over-decomposition.
• Add a reranker after the tool calls to filter noisy search results before feeding them to the sub-agent.
• Chain with compression if the synthesis prompt grows beyond the model's context window.

---

Related pages

• Planner — lower-level task decomposition.
• Orchestration — multi-agent pipeline patterns.
• Compression — token budget control.