Orchestrate comprehensive TODAS research for novel/emerging domains (1-7 subagents adaptive). Specializes in unprecedented topics, post-training data, and emerging technologies. Uses adaptive depth-based methodology: straightforward queries (1 agent), standard queries (2-3 agents), complex queries (5-7 agents). Handles depth-first (multiple perspectives), breadth-first (distinct sub-topics), and straightforward investigations. Triggers include "novel", "emerging", "2025", "2026", "unprecedented", "new technology", research on topics that didn't exist during training cutoff. Use when this capability is needed.
Orchestrate comprehensive TODAS research for novel/emerging domains (1-7 subagents adaptive). Specializes in unprecedented topics, post-training data, and emerging technologies. Uses adaptive depth-based methodology: straightforward queries (1 agent), standard queries (2-3 agents), complex queries (5-7 agents). Handles depth-first (multiple perspectives), breadth-first (distinct sub-topics), and straightforward investigations. Triggers include "novel", "emerging", "2025", "2026", "unprecedented", "new technology", research on topics that didn't exist during training cutoff. Use when this capability is needed.
Novel Domain Research Orchestration (Tier 5 TODAS)
What is TODAS: Tactical Optimization & Depth-Adaptive System - an adaptive research methodology that adjusts agent count (1-7) and research depth based on query complexity and novelty. Optimized for emerging domains and post-training information.
Quick Start
Analyze query novelty: Assess if topic is novel/emerging (post-training, unprecedented, 2025+ developments)
Identify most critical sub-questions (focus on essential, avoid every angle)
Prioritize by importance and expected complexity
Define clear boundaries between sub-topics (prevent overlap)
Plan aggregation strategy
Example: "Compare EU country tax systems" → retrieve EU countries list, define comparison metrics, batch research by region (Northern, Western, Eastern, Southern Europe)
For Straightforward Queries:
Identify most direct, efficient path to answer
Determine if basic fact-finding or minor analysis needed
Specify exact data points required
Determine most relevant sources
Plan basic verification methods
Create extremely clear task description for subagent
For all query types, evaluate each step:
Can this be broken into independent subtasks? (efficiency)
Would multiple perspectives benefit this? (depth)
What specific output is expected? (clarity)
Is this strictly necessary to answer query? (focus)
Output: Concrete research plan with clear subagent allocation.
Output: Complexity score and allocation recommendation for each dimension.
Phase 3b: Specialist Selection with Self-Challenge (CRITICAL QUALITY GATE)
For EACH dimension, perform rigorous specialist selection with adversarial validation:
Step 1: Requirements Analysis & Initial Selection
Dimension [N]: [Dimension Name]
Requirements Analysis:
What information is needed? (facts, trends, papers, market data, standards)
What sources are required? (academic journals, industry blogs, official docs, regulations)
Current vs future focus? (2024 current state vs 2025+ emerging trends)
Theoretical vs production focus? (research papers vs real-world implementations)
Candidate Specialists (list 2-3 types that COULD handle this dimension):
Candidate A: [specialist_type] - Strengths: [what they excel at for this dimension]
Candidate B: [specialist_type] - Strengths: [what they excel at for this dimension]
Candidate C: [specialist_type] - Strengths: [what they excel at for this dimension]
Initial Selection: [specialist_type]
Initial Rationale: [Why this specialist best matches the dimension requirements - be specific about the match between requirements and specialist capabilities]
Step 2: Self-Challenge Phase (🚨 MANDATORY - DO NOT SKIP)
The self-challenge phase prevents lazy defaulting and ensures optimal specialist matching.
🚨 CHALLENGE: "Wait, why didn't I choose [alternative_specialist] instead of [initial_selection]?"
For EACH alternative candidate (repeat 2-3 times per dimension):
Alternative: [alternative_specialist_type]
Gains if chosen:
What unique value would this specialist provide?
What perspectives/sources/capabilities does it have that initial selection lacks?
What dimension requirements would it serve BETTER?
Example: "trend-analyst would provide emerging 2025 patterns and weak signal detection that web-researcher doesn't offer"
Cons if chosen:
What would this specialist LACK compared to initial selection?
What dimension requirements would be UNDERSERVED?
What trade-offs would we accept?
Example: "trend-analyst forecasts FUTURE trends but dimension needs CURRENT production implementations"
Comparison: [initial_selection] vs [alternative]
Where initial wins: [Specific requirements where initial is stronger]
Where alternative wins: [Specific requirements where alternative is stronger]
Net assessment: [Which better matches the dimension's PRIMARY requirements?]
🚨 Critical question: Does this comparison reveal initial selection was suboptimal?
Verdict:
✅ KEEP [initial_selection] - Rationale: [Why initial still best after challenge]
OR
❌ SWITCH to [alternative] - Rationale: [Why alternative is actually better - challenge caught a mismatch]
Repeat challenge for Alternative B, Alternative C
Step 3: Final Selection Documentation
Dimension [N] - FINAL SELECTION: [specialist_type]
Final Rationale (after surviving self-challenge):
Chosen because: [Strengths that best match dimension requirements]
Alternatives considered and rejected:
[Alternative A]: Rejected because [specific weakness or mismatch for THIS dimension]
[Alternative B]: Rejected because [specific weakness or mismatch for THIS dimension]
Decision confidence: HIGH (explicit adversarial challenge performed and passed)
Example Self-Challenge Workflow
Dimension 5: Real-Time Coordination Mechanisms
Requirements:
WebSocket/SSE/WebRTC protocol documentation
Shopify Mobile Bridge architecture (current)
W3C MiniApp standardization status
Novel aspect: Query mentions "emerging in 2025" (cutting-edge focus)
Candidates:
web-researcher: Current documentation, engineering blogs, official specs
Net assessment: This dimension has BOTH current (Shopify, W3C status) AND future ("emerging 2025") aspects
🚨 WAIT: The "emerging in 2025" and "cutting-edge" keywords suggest future focus is PRIMARY!
Verdict: ❌ SWITCH to trend-analyst
Justification: Self-challenge revealed the "emerging 2025" and "cutting-edge" keywords indicate this is a FUTURE-focused dimension. trend-analyst's forecasting strength better matches the PRIMARY requirement (emerging patterns) than web-researcher's current documentation strength. Initial selection was suboptimal - self-challenge caught this mismatch.
Final Selection: trend-analyst ✅ (REVISED from web-researcher)
Allocation Justification: Simple dimension (score 3) with production focus. web-researcher provides 85% coverage. Academic gap irrelevant. Single specialist efficient and sufficient.
Output: Allocation plan with specialist count per dimension (most stay at 1, critical dimensions upgrade to 2-3).
Phase 3d: Budget Optimization & Final Verification
After all allocation decisions complete, perform final budget validation and repetition challenge:
🔴 Overrun (>10): MUST optimize or provide exceptional justification
Repetition Challenge (Quality Gate for Lazy Defaults)
🚨 CHALLENGE: "Did I default to [specialist] out of laziness rather than intentional strategy?"
If ANY specialist type used MORE than 1 time, re-examine EACH usage:
Specialist Type: [type used multiple times, e.g., "web-researcher"]
Used For: Dimension [A], Dimension [B], Dimension [C]
For EACH Dimension:
Dimension [A]: [Dimension Name]
Why this specialist?: [Original rationale from Phase 3b]
🚨 Repetition Challenge: "Did I choose web-researcher because it's GENUINELY optimal, or because I already chose it for Dimension B and defaulted to familiarity?"
Fresh Comparison:
web-researcher provides: [specific value for THIS dimension]
Alternative ([other specialist]) would provide: [what alternative offers]
Net assessment: Which BETTER matches Dimension A's PRIMARY requirements?
Verdict:
✅ REPETITION JUSTIFIED: [Why web-researcher genuinely optimal for THIS specific dimension, independent of other dimensions]
❌ LAZY DEFAULT DETECTED: Switch to [alternative] - [Why alternative actually better match]
Web research tasks: Deploy subagent (orchestrator delegates, not executes)
Challenging steps: Deploy additional subagents for more perspectives
Compare results: Use ensemble approach and critical reasoning
Throughout Execution:
Monitor progress: Continuously check if query being answered
Update plan: Adapt based on findings from subagents
Bayesian reasoning: Update priors based on new information
Adjust depth: If running out of time or diminishing returns, stop spawning and synthesize
Tactical optimization: Efficiency over completeness when appropriate
Output: Complete research findings from all subagents ready for synthesis.
TodoWrite Integration
Use TodoWrite to track research progress:
Before starting research:
TodoWrite([
{content: "Analyze query and determine type (Phase 1-2)", status: "in_progress", activeForm: "Analyzing query type"},
{content: "Assess dimension complexity scores (Phase 3a)", status: "pending", activeForm: "Assessing dimension complexity"},
{content: "Execute self-challenge phase for specialist selection (Phase 3b)", status: "pending", activeForm: "Executing self-challenge phase"},
{content: "Determine resource allocation counts per dimension (Phase 3c)", status: "pending", activeForm: "Determining resource allocation"},
{content: "Perform budget optimization and repetition challenge (Phase 3d)", status: "pending", activeForm: "Performing budget optimization"},
{content: "Log allocation decisions to project_logs/ (Phase 3e)", status: "pending", activeForm: "Logging allocation decisions"},
{content: "Spawn specialist agents based on final allocation (Phase 4)", status: "pending", activeForm: "Spawning specialist agents"},
{content: "Spawn fact-checker for critical dimension verification (Phase 4)", status: "pending", activeForm: "Spawning fact-checker"},
{content: "Synthesize findings from all specialists (Phase 5)", status: "pending", activeForm": "Synthesizing findings"},
{content: "Report completion with attribution and novelty assessment (Phase 6)", status: "pending", activeForm": "Reporting completion"}
])
As you progress, mark tasks completed and update status. The new Phase 3a-3e steps ensure quality decisions through:
- Complexity assessment (avoid under/over-allocation)
- Self-challenge (catch suboptimal selections)
- Resource allocation (justify specialist counts)
- Repetition challenge (prevent lazy defaults)
- Decision logging (traceability and validation)
Benefits:
User visibility into research progress
Clear phase tracking
Helps avoid skipping steps
Specialist Agent Selection (CRITICAL)
🚨 DO NOT use research-subagent - This is a generic worker type lacking specialized capabilities. You MUST use specialist agents based on research needs.
Available Specialist Agents:
Agent Type
Use When
Specialized Capabilities
web-researcher
General web queries, current information, broad topics
WebSearch, WebFetch, comprehensive web coverage
academic-researcher
Scholarly papers, research publications, scientific topics
Agent Registry Location: .claude/agents/ directory contains all specialist agent definitions.
Subagent Count Guidelines (Adaptive)
TODAS adjusts agent count based on complexity:
Query Complexity
Subagent Count
Example
Straightforward
1 specialist
"What is tax deadline this year?" → 1 web-researcher
Standard
2-3 specialists
"Compare top 3 cloud providers" → 3 web-researchers (one per provider)
Medium
3-5 specialists
"Analyze AI impact on healthcare" → 4 agents (academic-researcher, market-researcher, trend-analyst, web-researcher)
High
5-7 specialists
"Fortune 500 CEOs birthplaces/ages" → 7 web-researchers in wave 1, then 7 more (sequential batching)
Claude Code parallel limit: Maximum 10 parallel tasks, cap at 7 for safety (hooks overhead).
For queries requiring >7 agents: Use sequential batching:
Spawn 7 specialist agents in parallel
Wait for completion
Spawn next 7 specialist agents
Repeat until coverage complete
Principle: Prefer fewer, more capable specialists over many narrow ones (reduces overhead).
Minimum: Always spawn at least 1 specialist agent for ANY research task (orchestrator delegates, not executes).
Verification Phase: After specialists complete, spawn fact-checker for critical domains (security, compliance, novel topics).
Task Tool Spawning Instructions
How to spawn specialist agent instances:
Step 1: Plan Research Dimensions & Select Specialists
Example: "WebRTC + Web3 convergence in 2025" → 3 dimensions:
WebRTC current state and 2025 developments → web-researcher (current info)
Web3 technologies and decentralization trends → trend-analyst (emerging tech)
Convergence patterns and integration architectures → academic-researcher (research patterns)
Step 2: Call Task Tool for EACH Dimension in ONE Message
🚨 CRITICAL: Use SPECIALIST agent types (web-researcher, academic-researcher, etc.), NOT research-subagent
Correct parallel spawning pattern:
[Call Task tool:]
subagent_type: "web-researcher" ← USE SPECIALIST TYPE
description: "Research WebRTC 2025 developments"
prompt: "Research WebRTC's current state and 2025 roadmap. Focus on:
- Latest WebRTC implementations (2025 features)
- New codec support and performance improvements
- Browser compatibility and adoption trends
Research Path: docs/research-sessions/{session_id}/
SESSION_ID: {session_id}
SPAWNED_BY: internet-research-orchestrator
INVOCATION_CONTEXT: subagent
Output Requirements:
- Save findings to Research Path above
- File naming: webrtc-2025-research-subagent-001.json
- Follow schema: .claude/skills/research/json-schemas/research-output-schema.json
- Size limit: 22K tokens or 20K characters"
[Call Task tool again in SAME message:]
subagent_type: "trend-analyst" ← SPECIALIST for emerging tech
description: "Research Web3 decentralization trends"
prompt: "Research Web3 technologies and decentralization in 2025. Focus on:
- Emerging Web3 protocols (2025)
- Decentralized infrastructure architectures
- Real-world adoption and use cases
[Same tracking parameters and output requirements]"
[Call Task tool again in SAME message:]
subagent_type: "academic-researcher" ← SPECIALIST for patterns/research
description: "Research WebRTC+Web3 convergence patterns"
prompt: "Research integration patterns between WebRTC and Web3. Focus on:
- Decentralized video streaming architectures
- P2P communication with blockchain integration
- Novel use cases emerging in 2025
[Same tracking parameters and output requirements]"
All Task calls in ONE message = parallel execution (Claude Code optimization).
✅ Use SPECIALIST agent types (web-researcher, academic-researcher, trend-analyst, etc.)
✅ Use Task tool (it's available to you)
✅ Call Task multiple times in ONE message (parallel spawning)
✅ Spawn specialists IMMEDIATELY after planning (efficiency)
✅ Spawn fact-checker for critical domains (security, compliance, novel topics)
✅ Pass researchPath to ALL agents (when provided to skill)
✅ Include tracking parameters (SPAWNED_BY, SESSION_ID, INVOCATION_CONTEXT)
Delegation Rules (CRITICAL)
Main Claude delegates ALL research to subagents:
Core Orchestration Principles
Orchestrator role: You coordinate and synthesize, NOT execute primary research
Spawn 1-7 research-subagent instances (adaptive based on complexity)
Provide each subagent with extremely detailed, specific instructions
Let subagents perform all web searches, fact-finding, and information gathering
Focus on planning, analyzing, integrating findings, identifying gaps
Verification requirement: After planning, count how many subagents you plan to spawn
If count = 0, revise plan immediately
Every query requires AT LEAST 1 research-subagent
Your value is orchestration and synthesis, not execution
ResearchPath coordination: When invoked with researchPath parameter
ALL subagents MUST save outputs to SAME researchPath
Pass researchPath to EVERY subagent you spawn
File coordination is MANDATORY for multi-agent research
Subagent Task Descriptions
Provide each subagent with:
Specific research objective: Ideally 1 core objective per subagent
Expected output format: List, report, answer, analysis, etc.
Background context: How subagent contributes to overall research plan
Key questions: What to answer as part of research
Starting points and sources: Define reliable information, list unreliable sources to avoid
Specific tools: WebSearch, WebFetch for internet information gathering
Scope boundaries: Prevent research drift (if needed)
Output requirements: When researchPath provided (see section above)
Validation: If all subagents follow instructions well, aggregate results should allow EXCELLENT answer to user query (complete, thorough, detailed, accurate).
Deployment Strategy
Priority and dependency:
Deploy most important subagents first
If tasks depend on results from specific task, create that blocking subagent first
Ensure sufficient coverage for comprehensive research
All substantial information gathering delegated to subagents
Avoid overlap:
Every subagent should have distinct, clearly separate tasks
Prevent replicating work unnecessarily
Avoid wasting resources on redundant research
Efficiency while waiting:
Analyze previous results
Update research plan
Reason about user's query and how to best answer it
Do NOT idle waiting for subagents
Response Instructions
Before Providing Final Answer
Review most recent facts compiled during research process
Reflect deeply: Can these facts answer query sufficiently?
Provide final answer in format best for user's query
When invoked with researchPath: Return summary of research completed and file locations (research skill handles synthesis)
When invoked standalone: Format final research report in Markdown with proper source attribution
Source Attribution (Tier 5 Critical)
For novel/emerging domains, transparency is essential:
Inline references: "According to [Source Name]", "Research from [Organization] shows..."
Sources section: At end of report, list all key sources consulted
Credibility: Builds trust and allows verification
Recency: Note publication dates (critical for 2025+ topics)
Novelty Assessment
Include in synthesis:
Novelty level: Is this truly emerging? Post-training? Unprecedented?
Confidence: How much information available? (Emerging topics = less data)
Gaps: What's unknown? What requires future research?
Verification challenges: Difficult to verify emerging claims (note limitations)
Examples
Example 1: Straightforward Novel Query (1 Subagent)
User query: "What is Claude Sonnet 4.5 and when was it released?"
Synthesis: After all 4 subagents return, Main Claude integrates findings, identifies trade-offs across approaches, assesses novelty and emerging trends, provides source-attributed analysis.
Example 3: Breadth-First Novel Query (5-7 Subagents)
User query: "Compare the top 5 decentralized video streaming platforms in 2025 across performance, cost, adoption, and technology stack."
Synthesis: After all 5 subagents return, Main Claude creates comparison table, identifies patterns across platforms, assesses maturity and adoption trends, provides source-attributed recommendations.
Example 4: Multi-Dimensional Query with Self-Challenge & Resource Allocation (5-7 Specialists)
User query: "Research emerging approaches for secure push notifications in multi-tenant mini-app platforms for 2025, covering security architecture, infrastructure scalability, real-time coordination, and mobile-native implementations."
Phase 1-2: Query Analysis & Dimension Breakdown
Query type: Breadth-first (4 distinct dimensions)
Novelty: Very high (2025 emerging approaches, post-training)
Dimensions identified:
