Deep Research Methods

Deep Research Methods

Methodology reference for rigorous AI-agent research. Complements deep-research (tool usage for firecrawl/exa) with research thinking — how to decompose questions, evaluate sources, synthesize findings, and avoid bias.

When to Use

• Planning a research strategy or approach
• Decomposing a broad question into researchable sub-questions
• Evaluating source credibility or handling conflicting sources
• Synthesizing findings from multiple sources
• Avoiding confirmation bias or echo chambers
• Setting up multi-agent research fan-outs
• Deciding when to stop researching

When NOT to Use

• Conducting actual research on a topic → use deep-research
• Writing or editing a finished document → use a writing skill
• Quick factual lookups where methodology guidance isn’t needed

Routing guide:

Quick Reference

Depth calibration

Source credibility hierarchy

1. Peer-reviewed papers (highest)
2. Official documentation
3. Practitioner post-mortems / production case studies
4. Technical blog posts by domain experts
5. General blog posts
6. Forum discussions
7. AI-generated content (lowest — treat as hypotheses to verify)

Research pipeline

Decompose → Plan queries → Fan-out search → Evaluate sources → Deep-read → Synthesize → Report

Stopping criteria

Stop when: (1) new sources repeat known information, (2) every sub-question has 2–3+ independent sources, (3) contradictions are identified even if unresolved, (4) the next search is unlikely to change conclusions.

Anti-pattern checklist

• Sources from multiple ecosystems (not just one vendor’s blog network)?
• Disconfirming evidence actively sought?
• Claims evaluated on evidence quality, not source prestige?
• Citation chains checked (do 5 “sources” trace to 1 original study)?
• Token budget allocated across sub-questions, not all on the first one?

Question Decomposition

Three planning approaches

1. Planning-only: Generate research tasks directly from the user question. Fast but brittle with ambiguous queries.
2. Intent-to-planning: Clarify user intent before generating queries. Ask: “What decision will this research inform?” to bound scope.
3. Unified intent-planning: Generate preliminary plans while engaging the user. Surfaces assumptions early.

Decomposition recipe

1. Identify the DECISION the research will inform (not just the topic)
2. Extract key concepts that need operationalization
3. Map sub-questions with dependency ordering:
   • Independent sub-questions → fan-out in parallel
   • Dependent sub-questions → sequence (answer A informs query B)
4. For each sub-question, identify:
   • What source type would authoritatively answer this?
   • Is this a fast-moving or stable domain?
   • What would a DISCONFIRMING answer look like?
5. Assign token/time budget proportional to sub-question importance

Query formulation strategies

The search-read-infer loop

The agent does NOT search once and reason once. The canonical loop:

SEARCH → READ → UPDATE mental model → SEARCH AGAIN with better questions
         ↑                                            |
         └────────────────────────────────────────────┘

Each iteration: (1) Act — produce search actions, (2) Observe — capture outcomes and assess gaps, (3) Optimize — update strategy, (4) Remember — persist key findings.

Source Evaluation

Recency weighting

Cross-referencing protocol

Minimum 5–8 authoritative sources before finalizing any conclusion.

1. Citation chain detection: When 5 “different sources” all cite the same study, you have 1 evidentiary point, not 5.
2. Ecosystem diversity: Seek sources from competing vendors, different regions, opposing viewpoints.
3. Temporal clustering: If all evidence comes from one time period, the conclusion may reflect a trend, not a durable truth.
4. Methodology check: A rigorous study by unknowns outweighs an unsupported claim by a famous researcher.

Source quality checklist

For each source, assess:

• Does it cite its own sources?
• Is the author identifiable with relevant credentials?
• Is the publication venue reputable for this domain?
• Does it acknowledge limitations or counterarguments?
• Is it selling something? (vendor content requires extra skepticism)
• When was it published relative to domain rate of change?
• Has it been cited by other credible sources?
• Does it provide reproducible methodology or just conclusions?

Handling conflicting sources

1. Check methodology — which source has better evidence backing?
2. Check recency — in fast-moving domains, newer may reflect evolved understanding
3. Check scope — are they actually answering the same question?
4. Preserve the contradiction — report both positions rather than forcing resolution
5. Flag confidence impact — contradictions lower overall claim confidence

Multi-Source Synthesis

The four analytical passes

1. Consensus detection — which claims appear in 3+ independent sources? These form high-confidence findings.
2. Contradiction mapping — where do sources directly disagree? Map: Source A claims X because [evidence]. Source B claims not-X because [different evidence].
3. Gap identification — what questions are implied but not addressed? Often the most valuable output.
4. Cross-source narrative — organize thematically with confidence levels attached to each claim.

Dual-perspective retrieval

Retrieve evidence using BOTH the original claim AND its negation. This captures supporting and contradicting evidence simultaneously.

Original query: "MCP improves agent security"
Negated query:  "MCP security vulnerabilities" / "MCP does not improve security"
→ Aggregate BOTH result sets before synthesis

Confidence levels

• High (3+ independent quality sources agree, no contradictions) — state as finding
• Medium (2 sources agree OR quality sources with minor caveats) — state with qualifier
• Low (single source OR contradicted) — flag as tentative/contested
• Speculative (no direct evidence, inferred from adjacent findings) — label explicitly

Thematic organization

Wrong: “Smith found X. Jones found Y. Chen found Z.” Right: “The evidence shows X [Smith 2024, Jones 2023]. However, this may not hold at scale [Chen 2022], and one study found the opposite under condition W [Park 2025].”

Subagent Research Patterns

Fan-Out (production standard)

One lead agent spawns 3–5 subagents in parallel. Each receives a structured brief:

subagent_brief:
  objective: "Answer: [specific sub-question]"
  output_format: "structured summary with citations"
  boundaries: "Do NOT research [adjacent topic]"
  token_budget: 8000
  quality_gate: "minimum 3 independent sources before concluding"

Critical rule: each subagent gets a bounded, purposeful brief — NOT a dump of the orchestrator’s full history.

Adversarial / Debate

Multiple agents reason independently, then argue toward convergence. Use architecturally diverse models — homogeneous agents become polarized rather than converging on truth.

Iterative Deepening (WARP)

Pass 1: Breadth-first survey (landscape, key sources, major positions)
Pass 2: Deep-dive on highest-uncertainty sub-questions
Pass 3: Fill remaining gaps, resolve contradictions where possible
Pass 4: Final synthesis with confidence calibration

Council Mode

Three phases: (1) classify research question by complexity, (2) 3+ diverse models generate independent assessments, (3) structured consensus synthesis identifying agreement, disagreement, and unique findings. Reduces hallucination rates by ~36%.

Practical ceiling

5–10 parallel agents before communication overhead exceeds value. Under fixed budgets, single-agent deep reading often beats multi-agent shallow reading.

Anti-Patterns

Confirmation bias

Detection: all found sources agree with the initial hypothesis; no disconfirming evidence appeared.

Fix: force negation queries; allocate at least 20% of queries to “X failures” / “X criticism.”

Source echo chambers

Detection: all sources from same ecosystem, same time period, or same citation network.

Fix: actively seek competing ecosystems, critical reviews, alternative approaches; check if multiple sources trace to the same original claim.

Sycophantic convergence (multi-agent)

Detection: agents uncritically adopt peer views in debate.

Fix: anonymize debate contributions; use architecturally heterogeneous models; enforce explicit evidence requirements — agents may only update beliefs when presented with NEW evidence.

Premature convergence

Detection: research stopped after 2–3 agreeing sources on a topic deserving thorough investigation.

Fix: define stopping criteria BEFORE beginning; require source diversity (different authors, publications, years); run the “What would change my mind?” test.

Scope creep

Detection: research expanding far beyond original question; token budget exhausted on tangents.

Fix: maintain original research map; only expand when tangent directly impacts a mapped sub-question; apply “Will this change the decision?” test.

Report Structure

Required sections

1. Executive summary (2–4 sentences): Research question, headline finding, confidence level.
2. Methodology disclosure: Sources consulted, search strategies, tools, depth, limitations.
3. Findings per sub-question: Organized by theme, not source. Each: claim + evidence + citations + confidence + caveats.
4. Evidence quality assessment: Rate sources; flag citation chains.
5. Consensus vs outlier claims: Separate multi-source consensus from single-source claims.
6. Knowledge gaps: Sub-questions inadequately answered, unresolved conflicts, thin evidence areas.
7. Source list with quality ratings: URL, access date, source type, brief quality note.

Confidence annotation format

[HIGH CONFIDENCE] Multi-agent fan-out improves research quality by 60–90% over single-agent
approaches when sub-questions are independent. (Sources: Anthropic 2025, arXiv:2508.12752)

[LOW CONFIDENCE] Council mode with 5+ models may hit diminishing returns at 7 models.
(Source: single benchmark, not replicated)

[SPECULATIVE] Progressive confidence estimation may replace post-hoc quality ratings.
(Inferred from arXiv:2604.05952 direction, no production validation yet)

Token Economics

Multi-agent research uses approximately 15x more tokens than single-agent chat. Pre-allocate budget proportional to sub-question importance. Reserve 15% as contingency for unexpected findings.

Diminishing returns signal: if the last 3 sources added zero new claims, you have likely reached saturation.

Decomposition Prompt Template

I need to research: [TOPIC]
Decision this will inform: [DECISION]

Decompose into sub-questions. For each:
1. State the sub-question precisely
2. Identify what source TYPE would answer it best
3. Note dependencies on other sub-questions
4. Suggest 2–3 initial search queries
5. Define what a DISCONFIRMING answer would look like

Sources

Core references (May 2026):

1. Deep Research Agents Survey — arXiv:2506.18096
2. Deep Research Survey — arXiv:2508.12752
3. Progressive Confidence Estimation — arXiv:2604.05952
4. Dual-Perspective Retrieval — arXiv:2602.18693
5. Multi-Agent Consistency — arXiv:2603.24481
6. Identity Bias in Multi-Agent Debate — arXiv:2510.07517
7. Council Mode: Mitigating Hallucination — arXiv:2604.02923
8. Anthropic Multi-Agent Research System