The Four Agentic Paradigms: Technical Deep Dive
Overview
This document provides detailed analysis of each paradigm's architecture, implementation patterns, and practical applications.
Paradigm 1: Latent Space Clinician (LSC)
Position in Taxonomy
- Knowledge Source: Implicit (parametric memory)
- Agency Objective: Event Cognition (understanding/synthesis)
Core Philosophy
"Trust the model's emergent reasoning capabilities to function like an experienced clinical assistant, providing insights."
The LSC treats the LLM's parameters as an internalized medical curriculum, leveraging vast knowledge encoded during pre-training to perform creative synthesis and clinical reasoning without external tools.
Architecture Overview
┌─────────────────────────────────────────────────────────────────┐
│ LATENT SPACE CLINICIAN │
├─────────────────────────────────────────────────────────────────┤
│ INPUT │
│ ├── Dialogue History (H_t) │
│ ├── Current Query (Q_t) │
│ └── Observations (O_t) │
├─────────────────────────────────────────────────────────────────┤
│ DECOMPOSITION │
│ ├── Prompt Engineering / Chain-of-Thought │
│ ├── Tree of Thoughts (possibility exploration) │
│ └── Structured Query Generation │
├─────────────────────────────────────────────────────────────────┤
│ PLANNING │
│ ├── Internal Deliberative Cycle │
│ │ └── Consistency Check → Adjust → Iterate │
│ └── Analogical Reasoning Cycle │
│ └── Plans → Record → Reference │
├─────────────────────────────────────────────────────────────────┤
│ MEMORY │
│ ├── Parametric: Prompts, Templates, Clinical Corpus │
│ │ └── Injection into Latent Space (Strategy + Context) │
│ └── Non-Parametric: Context Extension, Recursive Summary │
│ └── Memory Proxy → Patient/Symptom Tracking │
├─────────────────────────────────────────────────────────────────┤
│ OUTPUT: Clinical Snapshot │
└─────────────────────────────────────────────────────────────────┘
Technical Components
1. Strategic Planning (Decomposition)
| Strategy | Description | Example Systems |
|---|---|---|
| Implicit Association | Probabilistic pattern matching from training | BioGPT, BioBART |
| Chain-of-Thought | Explicit reasoning chains mimicking diagnostic logic | Med-PaLM, Chain of Diagnosis |
| Causal Alignment | Decomposition based on domain-specific causal mechanisms | MedAgents |
| Multi-Perspective | Breaking queries into specialist viewpoints | MDagents |
2. Strategic Planning (Iteration)
- Self-Consistency: Generate multiple reasoning paths, aggregate via voting
- Counterfactual Reasoning: Challenge hypotheses by altering assumptions
- Predictive Validation: Update representations after each turn for coherence
3. Memory Management
| Type | Function | Techniques |
|---|---|---|
| Parametric | Internalized medical curriculum | In-context learning, Instruction fine-tuning |
| Non-Parametric | Patient context window | Long-context models (Clinical ModernBERT), Recursive summarization |
4. Collaboration
- Single-Agent: General practitioner model; efficient but cognitively limited
- Multi-Agent (Dominant): Central orchestrator synthesizes specialist outputs
- Multi-Agent (Distributed): Peer-to-peer debate without central controller
5. Evolution
- Continual Learning: Regularization to prevent catastrophic forgetting
- Experience Replay: Periodically retrain on successful diagnoses
- Model Expansion: Allocate new capacity for emerging medical knowledge
Strengths & Weaknesses
| Strengths | Weaknesses |
|---|---|
| Zero-shot reasoning on novel cases | Hallucination risk |
| Pattern recognition from vast training | Knowledge staleness |
| Creative synthesis | Opaque reasoning process |
| No external dependency latency | Cannot verify against current sources |
Best Use Cases
- Medical education and tutoring
- Preliminary clinical inquiry
- Patient-friendly report summarization
- Hypothesis generation
Paradigm 2: Grounded Synthesizer (GS)
Position in Taxonomy
- Knowledge Source: Explicit (external verification)
- Agency Objective: Event Cognition (understanding/synthesis)
Core Philosophy
"LLMs should function as powerful natural language interfaces to reliable external information rather than as knowledge creators."
The GS operates under strict explicit knowledge grounding, retrieving, integrating, and citing verifiable external evidence to construct an accurate clinical snapshot.
Architecture Overview
┌─────────────────────────────────────────────────────────────────┐
│ GROUNDED SYNTHESIZER │
├─────────────────────────────────────────────────────────────────┤
│ INPUT │
│ ├── Dialogue History (H_t) │
│ ├── Current Query (Q_t) │
│ └── Observations (O_t) │
├─────────────────────────────────────────────────────────────────┤
│ DECOMPOSITION │
│ ├── Sub-query Generation (q_1 ... q_K) │
│ └── Query Dependency Graph Construction │
├─────────────────────────────────────────────────────────────────┤
│ ACTION EXECUTION │
│ ├── Knowledge-Based: Structured DB/KG queries │
│ ├── Search Engine: Forage → Constrain → Attribute │
│ └── Tool-Use: Deterministic calculators, risk scores │
├─────────────────────────────────────────────────────────────────┤
│ MEMORY │
│ ├── Parametric: Semantic translator, Query generator │
│ └── Non-Parametric: Evidential ledger (firewalled context) │
│ ├── Noise Filtering │
│ ├── Compression │
│ └── Conflict Detection │
├─────────────────────────────────────────────────────────────────┤
│ ITERATION │
│ ├── Cyclical Gap-Filling (schema completion) │
│ ├── Evidence Hierarchy Resolution (conflict handling) │
│ └── Progressive Deepening (granularity refinement) │
├─────────────────────────────────────────────────────────────────┤
│ OUTPUT: Grounded Clinical Snapshot with Citations │
└─────────────────────────────────────────────────────────────────┘
Technical Components
1. Strategic Planning (Decomposition)
| Strategy | Description |
|---|---|
| Logical Inquiry Pathway | Map queries to verified clinical workflows |
| Multi-hop Dependency Graph | Orchestrate sequential/concurrent retrieval |
| Data-Driven Triggers | Dynamic sub-query generation based on findings |
2. Action Execution (Unique to Explicit Paradigms)
Knowledge-Based Actions:
- Formal queries over medical knowledge graphs
- Axiomatic logical operations with audit trails
- Schema-enforced semantic constraints
Search Engine Actions:
1. FORAGE: Multi-round strategic query sequence
2. CONSTRAIN: Firewall retrieved content from model parameters
3. ATTRIBUTE: Mandate fine-grained citation pointers
Tool-Use Actions:
- Drug dosage calculators
- Clinical risk stratification scores
- Pharmacokinetic simulation engines
3. Memory Management
| Type | Function |
|---|---|
| Parametric | Semantic translation, Query generation, Fluent synthesis |
| Non-Parametric | Evidential ledger with strict compartmentalization |
Critical Design: Information compartmentalization prevents "knowledge contamination" where verified external facts blend with potentially hallucinated internal knowledge.
4. Collaboration
- Asymmetric human-agent partnership (cognitive augmentation)
- GS acts as verifiable information provider, not autonomous decision-maker
- Multi-agent: Evidence aggregation hub or data flow pipeline
Strengths & Weaknesses
| Strengths | Weaknesses |
|---|---|
| Verifiable, auditable outputs | Limited to retrieved knowledge |
| Source-traceable reasoning | Retrieval bottlenecks |
| Mitigates hallucination | Multi-hop reasoning latency |
| High trustworthiness | Dependent on external source currency |
Best Use Cases
- EHR query interfaces
- Clinical guideline lookup
- Administrative data analysis
- Regulatory-compliant decision support
Paradigm 3: Emergent Planner (EP)
Position in Taxonomy
- Knowledge Source: Implicit (parametric memory)
- Agency Objective: Goal Execution (autonomous action)
Core Philosophy
"Grant the LLM high autonomy to dynamically devise multi-step plans to achieve complex clinical goals based on internal procedural intuition."
The EP operates as an active digital collaborator that autonomously shapes workflow trajectories through self-directed planning and execution.
Architecture Overview
┌─────────────────────────────────────────────────────────────────┐
│ EMERGENT PLANNER │
├─────────────────────────────────────────────────────────────────┤
│ INPUT │
│ ├── Dialogue History (H_t) │
│ ├── Current Query (Q_t) │
│ └── Observations (O_t) │
├─────────────────────────────────────────────────────────────────┤
│ DECOMPOSITION │
│ ├── Implicit Cognitive Frameworks (NLU → Analysis → Diagnosis) │
│ ├── Cognitive Map Construction │
│ └── Self-Consistency Path Generation │
├─────────────────────────────────────────────────────────────────┤
│ PLANNING │
│ ├── Tree-Planner (probabilistic action space) │
│ └── Follow Queries → Specific Actions │
├─────────────────────────────────────────────────────────────────┤
│ ITERATION │
│ ├── Self-Reflection and Critique │
│ ├── Feedback from Interaction History │
│ └── Dynamic Plan Adjustment │
├─────────────────────────────────────────────────────────────────┤
│ MEMORY │
│ ├── Parametric: Workflow patterns, Procedural intuition │
│ │ └── Meta-Knowledge interaction and evaluation │
│ └── Non-Parametric: Dynamic log, Status highlight │
│ └── External interface (Supervise/Control) │
├─────────────────────────────────────────────────────────────────┤
│ OUTPUT: Autonomous Workflow Execution │
└─────────────────────────────────────────────────────────────────┘
Technical Components
1. Strategic Planning (Decomposition)
| Strategy | Description |
|---|---|
| Self-Consistency | Generate diverse reasoning paths, aggregate via voting |
| Tree-Planner | Construct cognitive map before execution |
| Cognitive Cycle | Internalize expert-level frameworks as implicit scaffolds |
| Graph Transformers | Capture relational structure of medical concepts |
2. Strategic Planning (Iteration)
- Self-Reflection: Internal adversarial process to identify flaws
- Reflexion Framework: Self-corrective capability for complex reasoning
- Dialogue-Driven Feedback: Adapt strategies based on patient responses
3. Memory Management
| Type | Function |
|---|---|
| Parametric | Central engine for autonomous planning; encodes procedural knowledge |
| Non-Parametric | Tactical workspace; dynamic log of workflow state |
Key Insight: Parametric memory encodes not just declarative knowledge but procedural knowledge — the sequential actions and decision logic underlying clinical practice.
4. Collaboration
- Single-Agent: Autonomous practitioner steering dialogue
- Multi-Agent (Dominant): Central coordinator for meta-planning, delegates to workers
- Multi-Agent (Distributed): Peer negotiation for conflict resolution
5. Evolution
- Outcome-Driven Learning: Reinforce decision paths based on clinical results
- Meta-Learning: Abstract generalizable heuristics from experience
- Proactive Self-Repair: Identify capability limits, trigger targeted learning
Strengths & Weaknesses
| Strengths | Weaknesses |
|---|---|
| Flexible multi-step planning | Action hallucination risk |
| Creative problem-solving | Potential for outdated procedures |
| Adapts to open-ended tasks | Difficult to audit |
| Autonomous workflow progression | May exceed safety boundaries |
Best Use Cases
- Open-ended clinical problem-solving
- Novel/undocumented clinical scenarios
- Health education facilitation
- Shared decision-making support
Paradigm 4: Verifiable Workflow Automator (VWA)
Position in Taxonomy
- Knowledge Source: Explicit (external verification)
- Agency Objective: Goal Execution (autonomous action)
Core Philosophy
"Agent autonomy is strictly constrained within pre-defined, verifiable clinical workflows or decision trees."
The VWA serves as an intelligent natural language interface that drives deterministic, protocol-based process engines, ensuring maximum safety and predictability.
Architecture Overview
┌─────────────────────────────────────────────────────────────────┐
│ VERIFIABLE WORKFLOW AUTOMATOR │
├─────────────────────────────────────────────────────────────────┤
│ INPUT │
│ ├── Dialogue History (H_t) │
│ ├── Current Query (Q_t) │
│ └── Observations (O_t) │
├─────────────────────────────────────────────────────────────────┤
│ DECOMPOSITION │
│ ├── High-Level Objective Mapping │
│ ├── PICO Structuring (Population, Intervention, Comparison) │
│ └── Sub-Objective Generation │
├─────────────────────────────────────────────────────────────────┤
│ PLANNING │
│ ├── Node Selection in Pre-Defined Workflow │
│ └── Graph-of-Thought Clinical Pathway Navigation │
├─────────────────────────────────────────────────────────────────┤
│ ACTION EXECUTION │
│ ├── Knowledge-Based: Vector DB queries, KG traversal │
│ ├── Search Engine: Query optimization + validation │
│ └── Tool-Use: SQL generation, clinical calculators │
├─────────────────────────────────────────────────────────────────┤
│ ITERATION │
│ ├── Closed-Loop Locate-and-Correct │
│ ├── RL-Based Path Policy Optimization │
│ └── Hierarchical Cyclic Execution │
├─────────────────────────────────────────────────────────────────┤
│ MEMORY │
│ ├── Parametric: Semantic processor, Intent translator │
│ └── Non-Parametric: Clinical axiom repository, State register │
│ ├── Static: Evidence triplets, EHR graphs │
│ └── Dynamic: Action history, Entity maps │
├─────────────────────────────────────────────────────────────────┤
│ OUTPUT: Protocol-Compliant Workflow Execution │
└─────────────────────────────────────────────────────────────────┘
Technical Components
1. Strategic Planning (Decomposition)
| Strategy | Description |
|---|---|
| Structured Mapping | Align patient state to verified clinical workflow |
| PICO Decomposition | Standardize queries into clinical elements |
| Graph-of-Thought | Navigate logical sub-tasks in decision graph |
| Safety Boundaries | Halt execution when outside workflow distribution |
2. Strategic Planning (Iteration)
- Locate-and-Correct: Feedback from tool retrieval for precise alignment
- RL Path Optimization: Learn to minimize diagnostic costs while maximizing accuracy
- Hierarchical Execution: Multi-stage collaborative graph with delegation
3. Action Execution
Knowledge-Based:
- RAG over vectorized EHRs (CardioTRAP, CLI-RAG)
- Medical knowledge graph traversal (MedRAG)
- Hierarchical diagnostic KG queries
Search Engine:
- Context-aware query generation (MeSH mapping)
- Programmatic validation with entropy metrics
- Mandatory transparency (traceable evidence chain)
Tool-Use:
- SQL generation for EHR databases
- Clinical calculator invocation
- Expert system delegation
4. Collaboration
- Single-Agent: Sequential checklist execution
- Multi-Agent (Dominant): Attending physician model with MAKAR/TeamMedAgents
- Multi-Agent (Distributed): Care pathway pipeline (DAG-based handoffs)
5. Evolution
- Workflow Component Refinement: Evolutionary algorithms for prompt/tool optimization
- Behavioral Adaptation: Learn collaborative strategies from outcome feedback
- Meta-Tool Learning: Optimize interaction patterns with external systems
Strengths & Weaknesses
| Strengths | Weaknesses |
|---|---|
| Maximum safety and predictability | Rigid; struggles with anomalies |
| High trustworthiness and traceability | Optimization paradox in team settings |
| Protocol compliance | Requires significant engineering investment |
| Auditable decision chains | Limited to pre-defined workflows |
Best Use Cases
- High-stakes clinical execution
- Standardizable care pathways
- Triage and screening systems
- Regulatory-compliant automation
Comparative Summary Table
| Dimension | LSC | GS | EP | VWA |
|---|---|---|---|---|
| Knowledge | Implicit | Explicit | Implicit | Explicit |
| Agency | Cognition | Cognition | Execution | Execution |
| Primary Strength | Creative synthesis | Verifiable output | Flexible planning | Safe automation |
| Primary Weakness | Hallucination | Retrieval limits | Action drift | Rigidity |
| Planning Style | Reasoning chains | Query decomposition | Cognitive maps | Protocol mapping |
| Memory Focus | Context extension | Evidence ledger | Workflow log | State register |
| Collaboration | Peer consultation | Evidence hub | Task negotiation | Care pipeline |
| Evolution | Continual learning | Strategy refinement | Meta-learning | Workflow tuning |
| Best For | Education, inquiry | Decision support | Novel problems | Standard workflows |
Selection Guidance
Choose LSC when:
- Exploring hypotheses with limited data
- Need creative pattern recognition
- Low-stakes advisory context
- Human expert available for validation
Choose GS when:
- Accuracy and verifiability are paramount
- Working with structured knowledge bases
- Need auditable decision trails
- Integration with existing data systems
Choose EP when:
- Tasks require multi-step autonomous action
- Situations are novel or undocumented
- Flexibility outweighs predictability needs
- Building adaptive patient education systems
Choose VWA when:
- High-stakes clinical execution required
- Standard protocols exist and should be followed
- Maximum auditability needed
- Regulatory compliance is essential