Multi-Agent Research Synthesis

A methodology for conducting comprehensive research using multiple AI research agents, with a CLI AI agent orchestrating phased synthesis of results (AI-skill).

---

Table of Contents

• The Problem
• The Solution
• What This Methodology Solves
• Division of Responsibilities
  • Actor 1: Designer
  • Actor 2: Research Agent
  • Actor 3: Synthesizer
• Tasks for the CLI Agent
• Methodology Principles
• Methodology
• How to Use
• License

---

The Problem

Different research agents produce results of varying volume and quality due to:

• different domain configurations
• hidden system instructions
• internal search, processing, and filtering mechanics
• access to different websites

Using a single SOTA agent doesn't guarantee quality results — it may miss important sources, have bias toward certain regions/languages, or simply "dig in the wrong direction."

The obvious solution: use multiple (3+) research agents and combine the results.

How to do this effectively?

• Combined output volume (500KB+) doesn't fit in context
• LLMs trying to "digest everything at once" get confused and lose details
• Manual synthesis takes days and still loses information
• No consistent, verifiable, or repeatable results

The Solution

Parallel launch of multiple research agents → strict task preparation methodology → normalization of raw data and references → phased synthesis through a "living scaffold" with embedded instructions → fully validatable final document → executive summary.

What This Methodology Solves

Deduplication: A significant portion of information overlaps between agents — the methodology extracts all value without duplicates.

Enrichment: Each agent finds something unique — the methodology helps identify, verify, and use these findings.

Contradiction Resolution: Some data conflicts between sources — the methodology requires either verification through primary sources or explicit notation of discrepancies.

Validatability: Any fact in the final document can be verified — there's source attribution and bibliography.

Time Savings: Research can be completed in 1-2 working days instead of weeks (if done by a human analyst).

Division of Responsibilities

The user manages three AI actors, each with their own specialization:

Actor 1: Designer (conversational model)

What it does:

• Transforms the customer's request into a structured task
• Defines scope, eligibility, tags, and scales
• Prepares card templates and output formats
• Establishes source policies and quality checklists

Result: Self-contained Deep Research task (TASK.md)

Actor 2: Research Agent (deep research agents)

What it does:

• Collection and validation of cases by specified criteria
• Cards by template with C-ID
• Sources with URL and S-ID in Appendix
• Observed Patterns (observations with attribution)

What it does NOT do:

• Executive Summary
• Conclusions and recommendations
• Cross-synthesis
• Data normalization

Result: Raw reports (50-200KB per agent)

Actor 3: Synthesizer (CLI agent with bash access)

What it does:

• Structure, tag, and reference normalization
• Case and source deduplication
• ID mapping between agents
• Cross-synthesis and contradiction resolution
• Additional checks and validations
• Summary and recommendations

Results:

• NORMALIZATION_REPORT.md (created in Phase 3, updated in Phases 4-7)
• _case_references.json
• _sources_references.json
• SYNTHESIS.md
• SUMMARY.md

Tasks for the CLI Agent (Actor 3)

The methodology is implemented through a sequence of tasks. Each task represents one type of cognitive work for the agent; details are in the corresponding phases.

#	Task	Phase	Human Involvement
1	Structural normalization	3
2	Reference normalization	4	✓ Decision on Phase 5
2a	Re-attribution (if needed)	5	(Optional)
3	Case mapping	6.3.1
4	Source mapping	6.3.2	✓ Reference validation
5	SYNTHESIS.md scaffold + quantitative analysis	6.2, 6.3.3
6	Card synthesis	6.5.1
7	Cross-synthesis	6.5.2	✓ Escalations, contradictions
8	Validation and finalization	7
9	Executive Summary	8

Human Involvement Points

After task 2 — Based on reference normalization results, decide if Phase 5 (re-attribution) is needed. Triggers: >30% unused references, many unattributed statements.

After task 4 — Fresh look at references before starting synthesis. Verify: deduplication is correct, mappings haven't lost data.

After task 7 — Resolution of escalations from cross-synthesis: unresolvable contradictions, cases losing sources, decisions on [Inference].

Methodology Principles

The methodology is not dogma, but a set of project instructions that have proven effective. They can and should be adapted to specific research needs.

1. Task quality determines result quality — tag legend, scope, eligibility, output format in TASK.md. Without a clear task, agents will produce garbage.

2. One type of cognitive work = one task — the agent doesn't get confused, results are predictable, easy to debug.

3. Normalization — structure, tags, references are brought to a unified format before synthesis begins ("cleaning up after agents"). Without this, synthesis is impossible.

4. Synthesis state management — scaffold with TODOs guides the agent between sessions, case and source references are built before card synthesis, "empty" WRK files indicate processing completeness.

5. Large file handling strategy — grep/search by keywords instead of "swallowing 500KB." Read in parts, write incrementally.

6. Agent comparison — document which agent is better for what, reuse knowledge in future research.

7. Three levels of verification — fetch → search by URL + context → human-in-the-loop. Escalate if automatic resolution fails.

8. Contradictions are explicitly documented — with attribution of both positions, critical ones go to verification or escalation.

9. Summary with focus — audience and purpose determine structure, what goes to the top, which details to preserve.

10. Separation of technical and semantic work — scripts for consistency checking (ID, format, counters). Everything else — read and verify semantically, regex doesn't work on analytics.

Methodology

Detailed description of all phases (1-8) with templates, checklists, and examples:

• 🇬🇧 METHODOLOGY.md — English version
• 🇷🇺 METHODOLOGY_RU.md — Russian version

How to Use

1. Read the README — understand the approach and actors
2. Study Phase 1 — task preparation determines everything else
3. Choose research agents — minimum 3 (OpenAI, Claude, Gemini recommended)
4. Follow the phases — use as project instructions for Actor 3
5. Adapt — the methodology is not dogma, adjust to your domain

Minimum set: TASK.md (task) → raw agent reports → SYNTHESIS.md → SUMMARY.md

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Copyright (c) 2026 Askold Romanov

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made.

Full license text: https://creativecommons.org/licenses/by/4.0/