Model Tiers Reference
Strategic model routing for optimal cost-performance balance
"Use LOW tier for exploration, escalate to MEDIUM for implementation, reserve HIGH for validation"
Model Tier System
Diverga routes agents to appropriate AI models (Opus, Sonnet, Haiku) based on task complexity, ensuring optimal cost-performance balance.
HIGH
13 agents- ▸Theory selection with VS methodology
- ▸Research design validation
- ▸Meta-analysis orchestration
- ▸Humanization transformation
MEDIUM
9 agents- ▸Literature search & screening
- ▸Data integrity validation
- ▸AI pattern detection
- ▸Ethics review
LOW
2 agents- ▸Humanization verification
- ▸Vector database construction
Tier Selection Matrix
Choose appropriate tier based on task complexity and research stage:
| Task Complexity | Research Stage | Recommended Tier | Reason |
|---|---|---|---|
| Critical decision | Early (design) | HIGH | Foundation decisions impact entire project |
| Standard analysis | Middle (data) | MEDIUM | Balanced rigor and cost-efficiency |
| Simple lookup | Any | LOW | Fast, reliable, cost-effective |
| Complex orchestration | Any | HIGH | Multi-agent coordination needed |
| Batch processing | Data collection | MEDIUM/LOW | Volume over individual quality |
| Validation gate | Quality check | HIGH | Critical quality assurance |
Cost Optimization Strategy
Maximize value by strategic tier routing:
Exploration → Implementation → Validation
Start with LOW tier agents for exploration (F5, I3), escalate to MEDIUM for implementation (B1, B2), reserve HIGH for validation (A1, A2)
Parallel LOW-tier Processing
Use LOW-tier agents for batch tasks (F5, I3) that can run in parallel without checkpoints
Checkpoint-Gated Escalation
Only escalate to HIGH tier when checkpoint requires human decision (🔴 REQUIRED)
Complete Agent-Tier Matrix
All 24 agents organized by tier and category:
HIGH Tier (13 agents)
| ID | Agent | Category | Purpose | Checkpoint |
|---|---|---|---|---|
| A1 | ResearchQuestionRefiner | A | FINER/PICO/SPIDER formulation | 🔴 CP_RESEARCH_DIRECTION |
| A2 | TheoryCritiqueArchitect | A | Theory selection with VS | 🔴 CP_THEORY_SELECTION |
| A5 | ParadigmWorldviewAdvisor | A | Paradigm guidance | 🔴 CP_PARADIGM_SELECTION |
| C1 | QuantitativeDesignSampling | C | RCTs, quasi-experimental | 🔴 CP_METHODOLOGY_APPROVAL |
| C2 | QualitativeDesign | C | Phenomenology, GT | 🔴 CP_METHODOLOGY_APPROVAL |
| C3 | MixedMethodsDesign | C | Sequential, convergent | 🔴 CP_METHODOLOGY_APPROVAL |
| C5 | MetaAnalysisMaster | C | Multi-gate orchestration | 🔴 CP_META_GATE |
| D4 | MeasurementInstrumentDeveloper | D | Scale construction | 🔴 CP_METHODOLOGY_APPROVAL |
| E1 | QuantitativeAnalysisCodeGen | E | Statistical analysis | 🟠 CP_ANALYSIS_PLAN |
| E2 | QualitativeCodingSpecialist | E | Thematic, GT coding | - |
| E3 | MixedMethodsIntegration | E | Joint displays | 🟠 CP_INTEGRATION_STRATEGY |
| G6 | AcademicStyleHumanizer | G | Transform AI patterns | 🟡 CP_HUMANIZATION_VERIFY |
| I0 | SRPipelineOrchestrator | I | PRISMA pipeline coordination | 🔴 SCH_* checkpoints |
MEDIUM Tier (9 agents)
| ID | Agent | Category | Purpose | Checkpoint |
|---|---|---|---|---|
| B1 | LiteratureScout | B | PRISMA workflows | - |
| B2 | EvidenceQualityAppraiser | B | RoB, GRADE | - |
| D2 | DataCollectionSpecialist | D | Interview protocols | - |
| G1 | JournalMatcher | G | Target journal selection | - |
| G2 | PublicationSpecialist | G | Academic writing | - |
| G5 | AcademicStyleAuditor | G | AI pattern detection | 🟠 CP_HUMANIZATION_REVIEW |
| I1 | PaperRetrievalAgent | I | Multi-database fetching | 🔴 SCH_DATABASE_SELECTION |
| I2 | ScreeningAssistant | I | AI-PRISMA screening | 🔴 SCH_SCREENING_CRITERIA |
| X1 | ResearchGuardian | X | Research integrity, ethics | - |
LOW Tier (2 agents)
| ID | Agent | Category | Purpose | Checkpoint |
|---|---|---|---|---|
| F5 | HumanizationVerifier | F | Verify transformation | - |
| I3 | RAGBuilder | I | Vector database | 🟠 SCH_RAG_READINESS |
Temperature Settings by Category
Temperature controls creativity vs. consistency:
Strategic with creativity for alternatives
Precision in evidence synthesis
Balance rigor with design creativity
Structured but adaptive protocols
Analytical precision required
Maximum consistency for validation
Creative communication
PRISMA compliance precision
Balanced integrity oversight
Manual Model Override
You can override default tier by explicitly passing model parameter to Task tool:
// Default: A4 uses sonnet (MEDIUM tier)
Task(subagent_type="diverga:a4", model="opus", prompt="...")
// Override to HIGH tier for critical ethics review