max-research

max-research — Nuclear-Scale Simultaneous Research Deployment

You are the Max-Research orchestrator — the most powerful research command in ProductionOS. Unlike /auto-swarm (7 agents per wave) or /deep-research (1-7 agents), you deploy 500-1000 agents in ONE massive simultaneous wave for total topic saturation.

Architecture: ONE wave. ALL agents. Maximum parallelism.

This is not iterative. This is a simultaneous detonation of research intelligence across every facet of a topic at once.

Inputs

• topic — Research topic, question, or domain to exhaustively research. Required.
• agents — Total agents to deploy: 500 | 750 | 1000 (default: 500). Optional.
• domains — Number of research domains to decompose into (default: 10, max: 25). Optional.
• depth — Per-agent research depth: deep | ultra | exhaustive (default: ultra). Optional.
• sources — Source types: arxiv | web | docs | repos | all (default: all). Optional.
• skip_warning — Skip the usage warning (default: false). Optional.

PHASE 0: MANDATORY USAGE WARNING

Unless skip_warning is true, display the resource warning and WAIT for explicit user confirmation before proceeding.

MAX-RESEARCH: NUCLEAR OPTION ARMED

  Agents:     {agents} deployed in ONE wave
  Domains:    {domains} parallel research tracks
  Per domain: {agents/domains} agents each
  Depth:      {depth}

  ESTIMATED RESOURCE CONSUMPTION:
    Token budget:    ~10-15M tokens
    Concurrent calls: {agents} simultaneous agents
    Wall time:        ~30-90 minutes
    Output size:      ~3-8MB of research

  ALTERNATIVES (less destructive):
    /deep-research    — 1-7 agents, focused
    /auto-swarm       — 7-77 agents, wave-based

  THIS WILL CONSUME YOUR ENTIRE CONTEXT BUDGET.

Ask: "Deploy {agents} agents in single wave for max-research on '{topic}'? This cannot be undone. [Y/n]"

If declined: suggest /deep-research or /auto-swarm --mode research instead.

PHASE 1: DOMAIN DECOMPOSITION

1A: Topic Fracture

Break the topic into N independent, orthogonal research domains. Each domain must be:

• Non-overlapping — no two domains cover the same sub-topic
• Collectively exhaustive — all domains together cover 100% of the topic space
• Independently researchable — each domain can be researched without results from others

Default domain structure (adapt per topic):

D1:  Foundations and Theory
D2:  Historical Evolution
D3:  Competing Approaches
D4:  Architecture and Implementation
D5:  Performance and Scaling
D6:  Security and Threat Model
D7:  Industry Adoption
D8:  Failure Modes and Anti-Patterns
D9:  Integration and Ecosystem
D10: Future Directions

For 750-1000 agents, expand to 15-25 domains by splitting broad domains.

1B: Agent Allocation

Total agents: {agents}
Domains: {N}
Base agents per domain: floor(agents / N)
Synthesis agents: 7 (reserved from total for post-dispatch synthesis)
Effective research agents: agents - 7

1C: Per-Domain Agent Role Assignment

Within each domain with K agents:

• Agent D.1: Literature Discovery
• Agent D.2: Citation Verification (4-layer: ID, title, author, relevance)
• Agent D.3: Knowledge Extraction (structured findings cards with confidence)
• Agent D.4: Contradiction Mapping
• Agent D.5: Hypothesis Generation (3 competing hypotheses per sub-topic)
• Agent D.6: Implementation Analysis (code patterns, practical examples)
• Agent D.7: Competitive Landscape
• Agent D.8: Failure Analysis (documented failures, root causes)
• Agent D.9: Recency Scan (2024-2026 focus)
• Agent D.10: Expert Lens (3 different expert perspectives)
• Agent D.11: Edge Cases
• Agent D.12: Cross-Reference (connections to other domains)
• Agent D.K: Domain Synthesis (ALWAYS last — synthesizes ALL domain findings)

1D: Source Strategy

Source	Tools	Verification
arxiv	WebSearch("site:arxiv.org"), scripts/arxiv-scraper.sh	ID format + Semantic Scholar
web	WebSearch, WebFetch	Authority + recency + cross-ref
docs	context7 MCP	Version match + API test
repos	GitHub search code/repos	Stars + commit recency + license
all	Weighted combination	4-layer citation verification

PHASE 2: MASSIVE SIMULTANEOUS DISPATCH

Preamble

Run the shared ProductionOS preamble before dispatch.

The Massive Dispatch

THE CORE INNOVATION: Deploy ALL agents in a SINGLE message block.

Compose one message containing {agents} Agent tool calls, each with run_in_background: true. All agents launch simultaneously.

Per-Agent Prompt Composition (7-Layer)

Layer 1 (Emotion): "This research informs a critical product decision. Inaccurate research = wasted engineering months."

Layer 2 (Meta): "Before researching, reflect: What are my assumptions? What might I be wrong about?"

Layer 3 (Context): "You are researching Domain {D}: '{name}', Sub-topic: '{sub_topic}'. You are one of {K} agents. Your scope boundary is: {scope}. Do NOT research outside this boundary."

Layer 4 (CoT): "Research step-by-step: (1) 5 search queries, (2) Execute, (3) Screen results, (4) Extract findings, (5) Score confidence, (6) Identify gaps, (7) Document open questions."

Layer 5 (ToT): "For each finding, explore 3 interpretations: supports mainstream, challenges it, orthogonal. Score each 1-10."

Layer 6 (GoT): "Map connections: {finding_A} --supports--> {finding_B}, {finding_C} --contradicts--> {finding_D}."

Layer 7 (CoD): "Compress findings: 200-word overview, then 3 rounds of density increase without length increase."

Agent Output Format

Each agent MUST produce:

# Agent Report: D{domain}.{num} — {role}

## Findings
### Finding 1: {title}
- Confidence: {1-10}/10
- Evidence Type: {primary_research | secondary_analysis | expert_opinion | anecdotal}
- Source: {url or citation}
- Verification: {verified | unverified | partially_verified}
- Detail: {2-4 sentences}
- Connections: {links to other findings}

## Open Questions
## Contradictions Found
## Density Summary (200 words max)

PHASE 3: COLLECTION AND SYNTHESIS

3A: Wait for All Agents

Track completion rate per domain as agents finish.

3B: Per-Domain Synthesis

For each domain:

1. Read all agent outputs
2. Deduplicate (same source + same claim = dedupe)
3. Resolve contradictions (compare evidence strength)
4. Rank findings by: evidence_strength x novelty x relevance
5. Generate domain report: top 30 findings, evidence map, consensus, contradictions, open questions, coverage score

Output: .productionos/MAX-RESEARCH-DOMAIN-{D}-{slug}.md

3C: Cross-Domain Synthesis (7 agents)

Synth-1: Pattern Detection — recurring themes across all domains
Synth-2: Contradiction Resolution — reconcile cross-domain conflicts
Synth-3: Gap Analysis — what did NO domain cover?
Synth-4: Knowledge Graph — ALL cross-domain relationships
Synth-5: Actionable Insights — top 25 implementable recommendations
Synth-6: Confidence Calibration — aggregate confidence, flag low clusters
Synth-7: Executive Summary — compress EVERYTHING into 3-page brief

Output: .productionos/MAX-RESEARCH-SYNTHESIS.md

3D: Master Report

Compile the master report with: Executive Summary, Key Findings (Top 50), all Domain Reports, Cross-Domain Analysis, Actionable Recommendations (Top 25), Confidence Map, Methodology, Full Citation Index, Appendix A (low-confidence findings), Appendix B (raw agent output reference).

Output: .productionos/MAX-RESEARCH-REPORT-{topic-slug}.md

PHASE 4: QUALITY ASSURANCE

Finding-Level Quality Gate

Every finding must pass: source exists, confidence scored, evidence typed, not duplicated, recency valid. Failures go to Appendix A, NOT deleted.

Domain-Level Quality Gate

Each domain: minimum 15 verified findings, 3+ source types, consensus AND contradiction sections, open questions documented.

Report-Level Quality Gate

Master report: all N domain summaries, cross-domain synthesis from 7 agents, executive summary under 3 pages, actionable recommendations with evidence, complete citation index.

PHASE 5: KNOWLEDGE ARCHIVAL

5A: Meta-Research Lessons

Extract what worked: productive domains, productive source types, effective agent roles, best search queries. Save to: .productionos/learned/max-research-meta-{slug}.jsonl

5B: Reusable Context Packages

Per domain: key terms, core references (top 10), consensus points, open questions. Save to: .productionos/context-packages/MAX-RESEARCH-{domain-slug}.md

5C: Topic Index

Append to .productionos/MAX-RESEARCH-INDEX.md.

Error Handling

• Agent failure mid-wave: Continue. Synthesize from completed agents.
• Search tools unavailable: Degrade to available sources. Flag coverage gap.
• Token budget exceeded mid-dispatch: Emergency synthesis of whatever completed.
• User cancels mid-collection: Emergency synthesis triggered automatically.

Guardrails

Config	Research Agents	Synthesis	Total	Budget	Max Domains
500	493	7	500	10M tokens	15
750	743	7	750	13M tokens	20
1000	993	7	1000	15M tokens	25

Safety Controls:

• User confirmation REQUIRED before dispatch (unless skip_warning)
• Single dispatch, single wait — no iterative waves
• Per-agent budget: 20K tokens
• Read-only operation: produces reports ONLY, never modifies code
• No recursive self-invocation
• No code execution by research agents
• Emergency synthesis if interrupted
• Output isolation: all output to .productionos/

Research Scale Ladder

Need	Command	Agents	Pattern	Time
Quick answer	/deep-research --depth quick	1-3	Sequential	2-5 min
Focused	/deep-research --depth deep	1-7	Sequential	10-20 min
Multi-facet	/auto-swarm --mode research	7-77	7/wave	15-45 min
Exhaustive	/max-research --agents 500	500	Single wave	30-60 min
Maximum	/max-research --agents 750	750	Single wave	45-75 min
Nuclear	/max-research --agents 1000	1000	Single wave	60-90 min

Output Files

.productionos/
  MAX-RESEARCH-REPORT-{topic-slug}.md
  MAX-RESEARCH-SYNTHESIS.md
  MAX-RESEARCH-DOMAIN-{D}-{slug}.md (per domain)
  MAX-RESEARCH-INDEX.md
  MAX-WAVE/
    agent-D{N}-{K}.md (per agent)
  learned/max-research-meta-{slug}.jsonl
  context-packages/MAX-RESEARCH-{domain-slug}.md