Researcher Hand

Autonomous deep researcher that conducts exhaustive investigations with cross-referencing and fact-checking

Overview

Researcher Hand is an AI-powered deep research agent that conducts thorough investigations, cross-references sources, fact-checks claims, and produces comprehensive structured reports.

Category: Productivity
Icon: 🧪

What It Does

1. Analyze Question

   Decompose complex questions into sub-questions and identify source types

2. Multi-Source Research

   Execute searches across web, academic papers, news, and specialized databases

3. Cross-Reference

   Verify claims across multiple independent sources

4. Fact-Check

   Check primary sources, known debunkings, and authoritative databases

5. Synthesize & Report

   Generate structured reports with citations and confidence levels

Configuration

Research Depth

Setting	Sources	Passes	Use Case
Quick	5-10	1 pass	Fast answers, straightforward questions
Thorough	20-30	Cross-referenced	Balanced depth (default)
Exhaustive	50+	Multi-pass, fact-checked	Critical decisions, academic-level research

Output Style

Style	Format	Length	Use Case
Brief	Executive summary	1-2 pages	Quick answers, busy executives
Detailed	Structured report	5-10 pages	Standard research (default)
Academic	Formal paper	Variable	Research papers, formal citations
Executive	Key findings + recommendations	2-3 pages	Decision-makers

Quality Controls

Setting	Description
Source Verification	Cross-check claims across multiple sources
Max Sources	10, 30, 50, unlimited
Auto Follow-Up	Research tangential questions discovered during investigation
Save Research Log	Detailed log of queries and source evaluations
Citation Style	inline_url, footnotes, academic_apa, numbered

Activation

Basic Setup

openfang hand activate researcher

Configure research settings:

openfang hand config researcher \
  --set research_depth="thorough" \
  --set output_style="detailed" \
  --set source_verification="true" \
  --set max_sources="30" \
  --set citation_style="inline_url"

Example Workflow

Depth: Thorough (20-30 sources, cross-referenced)
Style: Detailed report
Verification: Enabled

> Research: What are the most effective AI agent architectures as of 2026?

Researcher Hand will:

1. Decompose into sub-questions:
   • What agent architectures exist?
   • How are they evaluated?
   • Which perform best on benchmarks?
   • What are real-world use cases?
2. Execute 15-20 targeted searches
3. Fetch and evaluate 25-30 sources
4. Cross-reference key claims
5. Fact-check critical assertions
6. Generate 8-page report with citations
7. Save as research_ai_agent_architectures_2026-03-06.md

How It Works

1. Question Analysis

Identifies question type and decomposes:

Type	Strategy	Example
Factual	Authoritative sources	”What is the capital of France?”
Comparative	Multi-perspective analysis	”React vs Vue in 2026?”
Causal	Evidence chains	”Why did Silicon Valley Bank fail?”
Predictive	Trend analysis	”Will AGI arrive by 2030?”
How-to	Step-by-step with examples	”How to build an AI agent?”
Survey	Comprehensive landscape	”What are all the LLM providers?”

Example decomposition:

Question: "What are the most effective AI agent architectures as of 2026?"

Type: Survey + Comparative

Sub-questions:
1. What agent architectures currently exist?
2. How is "effectiveness" measured in agent benchmarks?
3. Which architectures perform best on standard benchmarks?
4. What are the trade-offs (speed, cost, reliability)?
5. What are real-world deployment examples?

2. Search Strategy Construction

For each sub-question, build 3-5 queries:

Direct queries:

"AI agent architectures"
"LLM agent frameworks explained"
"agent architecture guide"

Expert queries:

"AI agent architecture research paper"
"site:arxiv.org agent architecture"
"agent framework comparison analysis"

Comparison queries:

"ReAct vs Plan-and-Execute agents"
"agent architecture pros cons"
"best AI agent framework 2026"

Temporal queries:

"AI agent architecture 2026"
"latest agent framework"
"agent architecture update"

Deep queries:

"agent architecture case study"
"agent benchmark data"
"agent architecture statistics"

3. Information Gathering

For each search query:

1. Search

   web_search(query) to find results

2. Evaluate

   Check URL domain and snippet relevance before fetching

3. Fetch

   web_fetch(url) for promising sources

4. Extract

   Key claims, data points, expert quotes, methodology, publication date, author credentials

Source quality evaluation (CRAAP test):

Criteria	Questions	Score
Currency	When published? Still relevant?	A-F
Relevance	Directly addresses question?	A-F
Authority	Who wrote it? What credentials?	A-F
Accuracy	Claims verifiable? Sources cited?	A-F
Purpose	Informational, persuasive, commercial?	A-F

Example evaluation:

Source: "Agent Architectures in 2026" - arxiv.org/abs/2601.12345
Currency: A (published Jan 2026)
Relevance: A (directly compares architectures)
Authority: A (researchers from Stanford, cited 45 times)
Accuracy: A (methodology described, datasets linked)
Purpose: A (academic research, no commercial bias)
Overall: A (authoritative source)

If save_research_log = true:

# Research Log: AI Agent Architectures

## Query 1: "AI agent architectures 2026"
Results: 10
Fetched: 3
- arxiv.org/abs/2601.12345 (A) - Comprehensive comparison
- medium.com/@author/agents (C) - Overview, lacks depth
- vendor.com/agents (D) - Commercial content, biased

4. Cross-Reference & Synthesis

If source_verification = true:

Verify key claims:

Claim: "ReAct agents outperform Plan-and-Execute on HotPotQA by 15%"

Sources:
1. arxiv.org/abs/2601.12345 - "ReAct: 68.2%, Plan-and-Execute: 59.1%"
2. paperswithcode.com/sota/hotpotqa - Confirms ReAct leads
3. github.com/react-paper - Official benchmark code

Verification: ✓ Verified (3 independent sources)
Confidence: High

Flag contradictions:

Claim: "LangGraph is the most popular agent framework"

Source A (blog): "LangGraph dominates with 50k+ GitHub stars"
Source B (GitHub): LangGraph has 12k stars, AutoGPT has 160k

Contradiction: ⚠️ Sources disagree
Resolution: Check primary source (GitHub actual stats)
Result: Source A is incorrect/outdated

Synthesis:

## Finding 1: Agent Architecture Landscape

Consensus view (5 sources agree):
- ReAct: Reasoning + acting in interleaved steps
- Plan-and-Execute: Separate planning and execution phases
- Reflection: Iterative self-critique and improvement
- LLM Compiler: Parallel tool execution with DAG planning

Minority view (1 source):
- "Hybrid architectures outperform pure approaches" (needs more evidence)

Gaps in knowledge:
- Limited data on production deployment costs
- No standardized benchmark for long-running agents

5. Fact-Check Pass

For critical claims:

1. Find primary source — Original research, official data, not secondary reporting
2. Check for debunkings — Search “[claim] debunked” or “[claim] false”
3. Verify statistics — Cross-check against authoritative databases
4. Flag weak evidence — Single-source claims, contested assertions

Confidence levels:

Level	Criteria
Verified	3+ authoritative sources confirm
Likely	2 sources or 1 authoritative source
Unverified	Single source, plausible but unconfirmed
Disputed	Sources disagree

6. Report Generation

Detailed Report (Default)

# Research Report: AI Agent Architectures in 2026
**Date**: 2026-03-06 | **Sources Consulted**: 28 | **Confidence**: High

## Executive Summary

As of March 2026, the AI agent architecture landscape has consolidated around
four primary approaches: ReAct (reasoning-acting), Plan-and-Execute (decomposition),
Reflection (self-critique), and LLM Compiler (parallelization). Empirical benchmarks
show ReAct leading on question-answering tasks (HotPotQA: 68%), while Plan-and-Execute
excels on multi-step workflows (WebShop: 71%). Production deployments favor hybrid
approaches combining strengths of multiple architectures.

## Detailed Findings

### 1. Agent Architecture Taxonomy

**ReAct (Reasoning + Acting)**  
Source: [ReAct: Synergizing Reasoning and Acting in Language Models](https://arxiv.org/abs/2210.03629)  
Confidence: Verified

ReAct interleaves reasoning steps ("Thought:") with actions ("Action:") in a loop.
Each cycle produces an observation that feeds the next reasoning step.

**Performance:**
- HotPotQA: 68.2% (verified, 3 sources)
- FEVER: 71.4% (verified, 2 sources)

**Strengths:** Interpretable, handles dynamic tasks  
**Weaknesses:** Sequential bottleneck, token-inefficient

---

### 2. Comparative Performance

| Architecture | HotPotQA | WebShop | GAIA | Avg |
|--------------|----------|---------|------|-----|
| ReAct | 68.2% | 59.1% | 34.2% | 53.8% |
| Plan-and-Execute | 59.1% | 71.3% | 41.0% | 57.1% |
| Reflection | 64.7% | 65.2% | 38.9% | 56.3% |
| LLM Compiler | 66.1% | 68.4% | 43.7% | 59.4% |

Source: [Agent Architecture Benchmark 2026](https://arxiv.org/abs/2601.12345)  
Confidence: High (primary research, peer-reviewed)

---

## Key Data Points

| Metric | Value | Source | Confidence |
|--------|-------|--------|------------|
| GitHub stars (LangGraph) | 12,034 | GitHub API | Verified |
| Median agent latency | 3.2s | Anthropic blog | Likely |
| Production cost (1M runs) | $850 | Estimate from sources | Unverified |

## Contradictions & Open Questions

**Contradiction:** Cost estimates vary widely  
Source A claims $500/1M runs, Source B claims $1200/1M.  
Likely explanation: Depends heavily on model choice and caching.

**Open question:** Long-running agent stability  
No standardized benchmark exists for agents running 100+ steps.  
Gap in research literature.

## Sources

### Primary Sources (A-tier)
1. [ReAct: Synergizing Reasoning and Acting](https://arxiv.org/abs/2210.03629) - Original ReAct paper, 1200+ citations
2. [Agent Architecture Benchmark 2026](https://arxiv.org/abs/2601.12345) - Comprehensive comparison, Stanford
3. [LLM Compiler: Parallel Function Calling](https://arxiv.org/abs/2312.04511) - UC Berkeley research

### Secondary Sources (B-tier)
4. [LangGraph Documentation](https://langchain.com/langgraph) - Official framework docs
5. [Anthropic: Building Reliable Agents](https://anthropic.com/blog/agents) - Engineering best practices

### Supporting Sources (C-tier)
6. [Medium: Agent Architectures Overview](https://medium.com/@author/agents) - Good overview, lacks rigor

[... 22 more sources listed with ratings ...]

Brief Report

# Research: AI Agent Architectures (2026)

## Key Findings

- **Four main architectures**: ReAct, Plan-and-Execute, Reflection, LLM Compiler
- **Best for QA**: ReAct (68% on HotPotQA)
- **Best for workflows**: Plan-and-Execute (71% on WebShop)
- **Best overall**: LLM Compiler (59% average across benchmarks)
- **Production trend**: Hybrid approaches combining multiple strategies

## Sources
1. [Agent Benchmark 2026](https://arxiv.org/abs/2601.12345) - Stanford research
2. [ReAct Paper](https://arxiv.org/abs/2210.03629) - Original framework
3. [Anthropic Agents Guide](https://anthropic.com/blog/agents) - Best practices
4. [LangGraph Docs](https://langchain.com/langgraph) - Implementation
5. [LLM Compiler](https://arxiv.org/abs/2312.04511) - Parallel execution

Academic Report

# A Survey of AI Agent Architectures in 2026

## Abstract

This survey examines the current landscape of AI agent architectures...

## Introduction

Autonomous AI agents have emerged as a critical application of large language
models (LLMs). This paper surveys the architectural approaches...

## Methodology

We conducted a systematic review of 28 sources including peer-reviewed papers,
official documentation, and benchmark repositories...

## Findings

### 3.1 ReAct Architecture

Yao et al. (2022) introduced ReAct, which synergizes reasoning and acting...

## Discussion

## Conclusion

## References

Anthropic. (2026). Building Reliable Agents. Retrieved from https://anthropic.com/blog/agents

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022).
ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629.

[... APA-formatted references ...]

Output

File	Description
research_[topic]_YYYY-MM-DD.md	Main research report
research_log_YYYY-MM-DD.md	Detailed query and source log (if enabled)

Dashboard Metrics

• Queries Solved — Research questions answered
• Sources Cited — Total unique sources used
• Reports Generated — Number of reports delivered
• Active Investigations — In-progress research

Tips & Best Practices

For best results:

• Use Thorough depth for most questions (Quick often misses nuances)
• Enable source verification for important decisions
• Review sources section — check if Hand found authoritative sources
• Ask follow-up questions if findings are unclear
• Export to PDF for sharing with stakeholders

Researcher Hand never fabricates sources or data. Every claim is traceable to an actual source. If you see “No reliable sources found,” that’s accurate — not a failure.

Common Issues

“No reliable sources found”
Question may be too niche, recent, or speculative. Try broadening the question.

“Sources disagree”
This is valuable information! The Hand will report both perspectives.

“Report is too long”
Switch to output_style="brief" for executive summaries.

“Missing key source”
If you know of a critical source, mention it: “Research X, and be sure to check Y source.”

Advanced Usage

Multi-Language Research

openfang hand config researcher --set language="spanish"

Investigar: ¿Cuáles son las mejores arquitecturas de agentes IA en 2026?

Follow-Up Research

Based on your previous research on agent architectures, now research:
Which architecture is best for production chatbots?

Comparative Deep-Dive

Research: Compare ReAct vs Plan-and-Execute agent architectures.
Depth: Exhaustive
Style: Academic

Next Steps

Collector Hand

Monitor research topics continuously

Predictor Hand

Make predictions based on research findings