Deep Research Methods

Parent: Deep Research · Published reference · snapshot 2026-09-08 · skill deep-research/references/deep-research-methods.md

↓ Facts as markdown all context files

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 finding

These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

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. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#deep-research-methods]

When to Use

Planning a research strategy or approach [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Decomposing a broad question into researchable sub-questions [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Evaluating source credibility or handling conflicting sources [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Synthesizing findings from multiple sources [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Avoiding confirmation bias or echo chambers [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Setting up multi-agent research fan-outs [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]
Deciding when to stop researching [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-to-use]

When NOT to Use

Conducting actual research on a topic → use deep-research [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-not-to-use]
Writing or editing a finished document → use a writing skill [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-not-to-use]
Quick factual lookups where methodology guidance isn't needed [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#when-not-to-use]

Source credibility hierarchy

Peer-reviewed papers (highest) [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-credibility-hierarchy]
Official documentation [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-credibility-hierarchy]
Practitioner post-mortems / production case studies [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-credibility-hierarchy]
Technical blog posts by domain experts [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-credibility-hierarchy]
AI-generated content (lowest - treat as hypotheses to verify) [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-credibility-hierarchy]

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. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#stopping-criteria]

Anti-pattern checklist

[ ] Sources from multiple ecosystems (not just one vendor's blog network)? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#anti-pattern-checklist]
[ ] Disconfirming evidence actively sought? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#anti-pattern-checklist]
[ ] Claims evaluated on evidence quality, not source prestige? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#anti-pattern-checklist]
[ ] Citation chains checked (do 5 "sources" trace to 1 original study)? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#anti-pattern-checklist]
[ ] Token budget allocated across sub-questions, not all on the first one? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#anti-pattern-checklist]

Three planning approaches

Planning-only: Generate research tasks directly from the user question. Fast but brittle with ambiguous queries. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#three-planning-approaches]
Intent-to-planning: Clarify user intent before generating queries. Ask: "What decision will this research inform?" to bound scope. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#three-planning-approaches]
Unified intent-planning: Generate preliminary plans while engaging the user. Surfaces assumptions early. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#three-planning-approaches]

Decomposition recipe

Identify the DECISION the research will inform (not just the topic) [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
Extract key concepts that need operationalization [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
Map sub-questions with dependency ordering: [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
- Independent sub-questions → fan-out in parallel [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
- Dependent sub-questions → sequence (answer A informs query B) [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
For each sub-question, identify: [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
- What source type would authoritatively answer this? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
- Is this a fast-moving or stable domain? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
- What would a DISCONFIRMING answer look like? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]
Assign token/time budget proportional to sub-question importance [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#decomposition-recipe]

The search-read-infer loop

The agent does NOT search once and reason once. The canonical loop: [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-search-read-infer-loop]
Each iteration: (1) Act - produce search actions, (2) Observe - capture outcomes and assess gaps, (3) Optimize - update strategy, (4) Remember - persist key findings. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-search-read-infer-loop]

Cross-referencing protocol

Minimum 5–8 authoritative sources before finalizing any conclusion. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#cross-referencing-protocol]
- Citation chain detection: When 5 "different sources" all cite the same study, you have 1 evidentiary point, not 5. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#cross-referencing-protocol]
- Ecosystem diversity: Seek sources from competing vendors, different regions, opposing viewpoints. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#cross-referencing-protocol]
- Temporal clustering: If all evidence comes from one time period, the conclusion may reflect a trend, not a durable truth. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#cross-referencing-protocol]
- Methodology check: A rigorous study by unknowns outweighs an unsupported claim by a famous researcher. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#cross-referencing-protocol]

Source quality checklist

For each source, assess: [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Does it cite its own sources? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Is the author identifiable with relevant credentials? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Is the publication venue reputable for this domain? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Does it acknowledge limitations or counterarguments? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Is it selling something? (vendor content requires extra skepticism) [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] When was it published relative to domain rate of change? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Has it been cited by other credible sources? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]
- [ ] Does it provide reproducible methodology or just conclusions? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-quality-checklist]

Handling conflicting sources

Check methodology - which source has better evidence backing? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#handling-conflicting-sources]
Check recency - in fast-moving domains, newer may reflect evolved understanding [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#handling-conflicting-sources]
Check scope - are they actually answering the same question? [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#handling-conflicting-sources]
Preserve the contradiction - report both positions rather than forcing resolution [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#handling-conflicting-sources]
Flag confidence impact - contradictions lower overall claim confidence [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#handling-conflicting-sources]

The four analytical passes

Consensus detection - which claims appear in 3+ independent sources? These form high-confidence findings. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-four-analytical-passes]
Contradiction mapping - where do sources directly disagree? Map: Source A claims X because [evidence]. Source B claims not-X because [different evidence]. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-four-analytical-passes]
Gap identification - what questions are implied but not addressed? Often the most valuable output. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-four-analytical-passes]
Cross-source narrative - organize thematically with confidence levels attached to each claim. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#the-four-analytical-passes]

Dual-perspective retrieval

Retrieve evidence using BOTH the original claim AND its negation. This captures supporting and contradicting evidence simultaneously. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#dual-perspective-retrieval]

Confidence levels

High (3+ independent quality sources agree, no contradictions) - state as finding [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confidence-levels]
Medium (2 sources agree OR quality sources with minor caveats) - state with qualifier [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confidence-levels]
Low (single source OR contradicted) - flag as tentative/contested [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confidence-levels]
Speculative (no direct evidence, inferred from adjacent findings) - label explicitly [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confidence-levels]

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]." [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#thematic-organization]

Fan-Out (production standard)

One lead agent spawns 3–5 subagents in parallel. Each receives a structured brief: [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#fan-out-production-standard]
Critical rule: each subagent gets a bounded, purposeful brief - NOT a dump of the orchestrator's full history. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#fan-out-production-standard]

Adversarial / Debate

Multiple agents reason independently, then argue toward convergence. Use architecturally diverse models - homogeneous agents become polarized rather than converging on truth. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#adversarial-debate]

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%. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#council-mode]

Practical ceiling

5–10 parallel agents before communication overhead exceeds value. Under fixed budgets, single-agent deep reading often beats multi-agent shallow reading. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#practical-ceiling]

Confirmation bias

Detection: all found sources agree with the initial hypothesis; no disconfirming evidence appeared. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confirmation-bias]
Fix: force negation queries; allocate at least 20% of queries to "X failures" / "X criticism." [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#confirmation-bias]

Source echo chambers

Detection: all sources from same ecosystem, same time period, or same citation network. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-echo-chambers]
Fix: actively seek competing ecosystems, critical reviews, alternative approaches; check if multiple sources trace to the same original claim. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#source-echo-chambers]

Sycophantic convergence (multi-agent)

Detection: agents uncritically adopt peer views in debate. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sycophantic-convergence-multi-agent]
Fix: anonymize debate contributions; use architecturally heterogeneous models; enforce explicit evidence requirements - agents may only update beliefs when presented with NEW evidence. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sycophantic-convergence-multi-agent]

Premature convergence

Detection: research stopped after 2–3 agreeing sources on a topic deserving thorough investigation. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#premature-convergence]
Fix: define stopping criteria BEFORE beginning; require source diversity (different authors, publications, years); run the "What would change my mind?" test. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#premature-convergence]

Scope creep

Detection: research expanding far beyond original question; token budget exhausted on tangents. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#scope-creep]
Fix: maintain original research map; only expand when tangent directly impacts a mapped sub-question; apply "Will this change the decision?" test. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#scope-creep]

Required sections

Executive summary (2–4 sentences): Research question, headline finding, confidence level. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Methodology disclosure: Sources consulted, search strategies, tools, depth, limitations. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Findings per sub-question: Organized by theme, not source. Each: claim + evidence + citations + confidence + caveats. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Evidence quality assessment: Rate sources; flag citation chains. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Consensus vs outlier claims: Separate multi-source consensus from single-source claims. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Knowledge gaps: Sub-questions inadequately answered, unresolved conflicts, thin evidence areas. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]
Source list with quality ratings: URL, access date, source type, brief quality note. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#required-sections]

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. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#token-economics]
Diminishing returns signal: if the last 3 sources added zero new claims, you have likely reached saturation. [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#token-economics]

Sources

Core references (May 2026): [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Deep Research Agents Survey - arXiv:2506.18096 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Deep Research Survey - arXiv:2508.12752 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Progressive Confidence Estimation - arXiv:2604.05952 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Dual-Perspective Retrieval - arXiv:2602.18693 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Multi-Agent Consistency - arXiv:2603.24481 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Identity Bias in Multi-Agent Debate - arXiv:2510.07517 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Council Mode: Mitigating Hallucination - arXiv:2604.02923 [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]
- Anthropic Multi-Agent Research System [archived source: https://llms-explorer.com/sources/mdb-context-hub/deep-research-methods/#sources]

Where this helps

Deciding how many independent sources are needed before treating a factual claim as confirmed, using the source credibility hierarchy instead of stopping at the first plausible-sounding page. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Recognizing when a research pass has hit diminishing returns and should stop, using explicit stopping criteria instead of running endless additional queries. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Structuring a genuinely contested question, where sources disagree, using the adversarial/debate or dual-perspective retrieval approach instead of averaging conflicting claims into a mushy middle. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Catching confirmation bias in your own research process, where you keep searching until you find a source that agrees with your initial hypothesis and stop there. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

How to apply this

Before starting a research task, write down explicit stopping criteria — such as "three independent primary sources agree" or "no new information in two consecutive searches" — so you don't quit too early or spin forever. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Run the four analytical passes as a discipline rather than a single read-through, deliberately re-reading sources for a different purpose each pass. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
When two credible sources conflict, don't average their claims into a hedge — note the disagreement explicitly, along with each source's confidence level and why they might differ. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Apply the source credibility hierarchy explicitly when weighing claims, treating a primary source such as official documentation or an original filing as stronger evidence than a secondary summary or aggregator site. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Common mistakes

Premature convergence — settling on an answer after the first search pass instead of running the cross-referencing protocol to check whether other sources actually agree. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Scope creep — letting a research question expand mid-investigation until the original question is never actually answered. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Sycophantic convergence in multi-agent research setups, where subagents converge on whichever narrative was found first rather than genuinely cross-checking it. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Treating a large volume of sources as equivalent to a large volume of independent evidence, without checking whether they're actually just citing each other (a source echo chamber). [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Known issues

Source echo chambers are a real risk — many secondary sources cite the same original claim, so five agreeing sources can still mean only one actual independent data point. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
The four-pass, dual-perspective methodology adds real overhead versus a single-pass search-and-summarize, which isn't worth it for low-stakes questions. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Confidence levels assigned to findings are inherently subjective; different researchers applying the same methodology to the same sources can land on different confidence ratings. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
There's a practical ceiling: even a rigorous multi-pass methodology can't produce more certainty than the underlying sources actually contain — it structures reasoning, it doesn't manufacture missing evidence. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Children

Question Decomposition (frontier)
Source Evaluation Heuristics (frontier)
Multi-Source Synthesis (frontier)
Subagent Research Patterns (frontier)
Anti-Pattern Detection (frontier)

Frontier under this node: Anti-Pattern Detection, Multi-Source Synthesis, Question Decomposition, Source Evaluation Heuristics, Subagent Research Patterns

← the whole tree · 3D view· how to read this page