The Filter Bubble Problem
Traditional research can sometimes fall short. When you rely on a single search engine or database, you may end up in a “filter bubble,” where algorithmic biases shape what you see. This can quietly limit exposure to important information and diverse perspectives, resulting in a narrower understanding of your topic.
That’s the problem we set out to solve with the Deep Research Agent.
A Multi-Source Approach
The solution: route your research through multiple AI providers, synthesize their findings, and compare evaluations from different AI judges. This delivers richer context and a more complete understanding than any single provider could offer alone. You make the final call — reviewing the analyses and selecting the insights that best serve your research goals.
Think of it like tuning into different media channels. When you want a fuller picture of a complex topic, you don’t rely on just one source — you switch between networks and local outlets. Each offers its own perspective, strengths, and blind spots. Comparing them brings you closer to the full story.
The Deep Research Agent works the same way.
How It Works: A Simple 4-Step Process
Step 1: Define Your Topic
Start with a clear research question. The more specific, the better. Think:
- “How does quantum computing threaten modern cryptography?”
- “What are sustainable urban planning strategies for coastal cities?”
Then pick your planning engine — the AI that will create your research blueprint. Maybe you choose Claude for its structured thinking, or perhaps Gemini for its breadth of knowledge. Hit Generate Research Plan.
Step 2: Review and Refine Your Plan
The AI generates a structured plan with objectives and subtopics. But here’s the thing: you’re still the expert. Edit the plan. Add missing angles, rephrase things, tweak the search queries. A few minutes here transforms good research into great research.
Step 3: Configure and Execute
Now comes the fun part — choose your engines:
- Search Engine: Which AI will hunt down the sources? Perplexity is great for web search, while others might excel at academic databases.
- Synthesis Engine: Which AI writes the final report? Claude’s prose is clean and structured; DeepSeek might bring a different voice. Set your target word count and click Generate New Research.
Step 4: Evaluate and Synthesize
Choose one or more AI judges to evaluate your reports against the original research plan, scoring metrics like objective fulfillment, question coverage, and citation quality.
Now pick your path:
Merge the best: Have an AI judge fuse your top 2+ reports, combining their strengths — better depth from one, stronger analysis from another — into a single enhanced version that surpasses any individual report
Use the winner: Select your highest-scoring report as-is
The Multi-Source Workflow
Once you have your first report, the real power kicks in. Here’s what makes the Deep Research Agent different:
Research with Different Providers
Hit Re-Synthesize and choose a different AI model. Maybe your first report used Perplexity for research and Claude for writing. Now try Google for research and OpenAI for writing. Same plan, different perspectives.
Compare Reports
With multiple reports in hand, the differences become clear. One might offer superior technical depth, another stronger analysis, a third richer examples. This comparison reveals each AI’s unique strengths.
Run Through Multiple Judges
This is where you can experiment: send your reports through different AI judges, each evaluating them against your original research plan. Although every scoring engine uses the same evaluation criteria, each AI model may interpret and weight them differently:
- Objective fulfillment
- Question coverage
- Evidence & citation quality
- Organization & coherence
- Depth & insight
By comparing scores across multiple AI judges, you gain a multi-perspective view of report quality, revealing which reports truly excel and why.
Pick the Best or Create Something Superior
You have several options:
- Select your top report based on the scoring and use them as-is
- Generate a synthesis: Choose a judge to merge the strongest elements from multiple reports, resolving contradictions and filling gaps
- Try another judge: Not satisfied? Run a different AI judge and see if it highlights different strengths
- Iterate again: Generate more reports with new engine combinations, then re-evaluate
The Fuse Advantage
The AI Judge analyzes multiple research reports and creates a unified synthesis that preserves the strongest elements from each source. The Judge is assisted by two analysis tools:
Draft Statistics – Shows for each report:
- Word count
- Number of citations
- Citation density (citations per 100 words)
- Percentage of paragraphs with citations
Citation Overlap Analysis – Reveals:
- Total unique sources across all reports
- Consensus sources (cited by 2+ reports) – widely-recognized information
- Unique sources (cited by only 1 report) – distinctive discoveries from specific search engines
- Which report found which unique sources
Core Fusion Approach
- Preserves unique insights – Uses overlap analysis to identify and value content found by only one search engine
- Combines complementary coverage – Merges different aspects covered by different reports
- Balances consensus with unique – Includes both widely-cited information (consensus) and distinctive discoveries (unique)
- Resolves contradictions – Presents multiple viewpoints with their respective citations
- Improves structure – Better organization, flow, and transitions
- Maintains citation density – Preserves source support throughout
Content Handling
- Prioritizes cited content – Information with citations is preserved preferentially
- Handles uncited content carefully – May include factual-seeming content without citations if important
- Consolidates redundancy – Overlapping information is stated once with citations from all sources
- Aims for comprehensiveness – The final report combines unique insights from all sources
Result: A coherent, well-structured report that’s more comprehensive than any single source. The judge adds value through integration and organization.
Loop Until Satisfied
The process is iterative. Generate reports, compare judges’ evaluations, synthesize, and repeat until you have exactly what you need. Mix and match engines and judges freely — each combination reveals something new.
Why This Matters
In a world of information overload and echo chambers, having a tool that deliberately seeks diverse perspectives is invaluable. Whether you’re a student, a professional researcher, or just someone who loves learning deeply, the Deep Research Agent turns what used to take hours into minutes while giving you a more balanced, comprehensive view.
Coming Soon: Open Source
We believe in the power of community-driven innovation. That’s why we’re planning to open source the Deep Research Agent soon. Developers will be able to:
- Add new search engines — Integrate additional web and academic search APIs
- Add new LLM providers — Extend support for more AI models as they emerge
- Devise advanced judging mechanisms
- Multi-Plan Comparative Analysis
- Compare reports against multiple research plans on same topic
- Identify coverage gaps: “What questions did Report A answer that Report B missed?”
- Detect contradictory conclusions and flag for resolution
- Specialized Judge Personas
- Peer Review Judge: Simulates academic peer review criteria
- Skeptic Judge: Actively looks for weak evidence and logical fallacies
- Policy Analyst Judge: Evaluates for decision-making suitability
- Technical Validator Judge: Checks code accuracy, technical claims
- Fact-Verification Layer
- Automated fact-checking against primary sources
- Citation chain validation (does source actually support the claim?)
- Statistical significance verification
- Multi-Plan Comparative Analysis
- Customize workflows — Modify the research pipeline to suit specific use cases
- Build on top of this foundation — Create new features and applications
Stay tuned!
Explore More
- See how it works: Deep Research
- See the workflow diagram: View Flow Chart
Deep Research Agent was developed by Dr. Ng Chong at UNU Campus Computing Centre. It’s open, modular, and designed to make deep, unbiased research accessible to everyone.
Ready to flip between some channels and see what turns up?