Agent System Types (A/B/C/D Taxonomy)

Core Concept: Taxonomy of AI agent systems based on their capabilities (data sources) Origin: Validated architecture from FluSisTip project (healthcare production system) Applicability: Generic for any regulated domain or rich-context requirements

🎯 Overview

In production AI agent development, especially in regulated domains (healthcare, legal, finance), not all agents need the same capabilities. The A/B/C/D typology classifies agents by type of data they access, enabling cost, performance, and complexity optimization.

Complexity Hierarchy

Type A (Pure AI)
  ↓ +Database Context
Type B (AI + CAG)
  ↓ +Web Grounding
Type C (AI + Web)
  ↓ +Database Context
Type D (AI + CAG + Web)

📊 Quick Comparison

Cost Legend:

• $ = $0.001 - $0.05
• $$ = $0.05 - $0.15
• $$$ = $0.15 - $0.30
• $$$$ = $0.30 - $0.50

🔷 Type A: Pure AI

Definition

Direct input → LLM → Structured output

Agent processes only data provided in input (by user or previous system in workflow). No database access, external APIs, or web search.

Characteristics

• ✅ Fastest: ~3s average time
• ✅ Cheapest: ~$0.02 per execution
• ✅ Deterministic: Same input = same output
• ✅ Stateless: No state between calls
• ❌ Limited context: Only what's in the prompt

When to Use

1. Analysis & Extraction: Identify entities, extract structured data
2. Classification: Categorize content, detect sentiment
3. Transformation: Convert formats, restructure data
4. Syntactic Validation: Check format, data completeness
5. Query Generation: Create search strategies

Healthcare Example: S.1.2 - Claims Identification

Function: Identify medical claims in raw text requiring scientific validation

Implementation Pattern (Clojure):

(ns lab.agents.claims-identifier
  "Type A: Pure AI - Medical claims identification"
  (:require [lab.config.google-cloud :as gc])
  (:import [com.google.cloud.vertexai VertexAI]
           [com.google.cloud.vertexai.generativeai GenerativeModel]))

(defn create-claims-agent
  "Creates Type A agent for medical claims identification."
  [{:keys [project-id location model]
    :or {model "gemini-1.5-flash"}}]
  (let [vertex-ai (VertexAI. project-id location)
        generation-config (-> (GenerationConfig/newBuilder)
                              (.setTemperature 0.2)  ; Low for technical analysis
                              (.setMaxOutputTokens 2000)
                              (.build))
        model-instance (.getGenerativeModel vertex-ai model)]
    (.withGenerationConfig model-instance generation-config)))

(defn identify-claims
  "Identifies medical claims in content."
  [agent content specialty]
  (let [prompt (format
                 "Analyze this %s content and extract ALL scientific claims
                 that require validation. Return JSON with 'claims' array."
                 specialty content)
        response (.generateContent agent prompt)
        text (-> response .getText)]
    (cheshire.core/parse-string text true)))

Real Metrics (Healthcare Pipeline):

• Average time: 3.2s
• Average cost: $0.018
• Success rate: 96.1%
• Average tokens: 520 total

🔶 Type B: AI + CAG (Context-Aware Generation)

Definition

Input + Database Context → LLM → Personalized output

Agent enriches prompt with data from tenant database (Context-Aware Generation). Enables personalization based on history, profile, stored business rules.

Characteristics

• ✅ Rich context: Access tenant data (regulations, profiles, history)
• ✅ Personalized: Adapts output to specific context
• ✅ Compliance: Incorporates stored business rules
• ⚠️ Query overhead: +2s to fetch contexts
• ⚠️ Cache important: Contexts should be cached (TTL 1h+)

When to Use

1. Output Personalization: Adapt tone, style, language to profile
2. Dynamic Compliance: Apply tenant-specific rules
3. History: Consider previous interactions/decisions
4. Custom Configuration: Client-specific guidelines
5. Rule Validation: Check conformance with stored policies

Healthcare Example: S.3-2 - SEO + Professional Profile

Function: Optimize content for SEO considering professional profile and specialized keywords

Implementation Pattern (Clojure):

(ns lab.agents.seo-optimizer
  "Type B: AI + CAG - SEO optimization with tenant context"
  (:require [lab.config.google-cloud :as gc]
            [next.jdbc :as jdbc]))

(defn fetch-professional-profile
  "Fetches complete professional profile (Context 1)."
  [db-spec prof-id]
  (jdbc/execute-one!
    (jdbc/get-datasource db-spec)
    ["SELECT nome_completo, crm, especialidade, anos_experiencia,
             tom_voz, cidade_atuacao
      FROM profissionais
      WHERE id = ?::uuid"
     (str prof-id)]
    {:builder-fn rs/as-unqualified-lower-maps}))

(defn fetch-seo-keywords
  "Fetches specialized SEO keywords (Context 2)."
  [db-spec specialty city]
  (jdbc/execute!
    (jdbc/get-datasource db-spec)
    ["SELECT keyword, tipo, search_volume
      FROM seo_keywords
      WHERE especialidade = ? AND cidade = ?
      ORDER BY relevancia DESC LIMIT 10"
     specialty city]
    {:builder-fn rs/as-unqualified-lower-maps}))

(defn optimize-seo
  "Optimizes content for SEO with professional context."
  [agent db-spec content prof-id]
  (let [profile (fetch-professional-profile db-spec prof-id)
        keywords (fetch-seo-keywords db-spec
                                     (:especialidade profile)
                                     (:cidade_atuacao profile))
        prompt (build-enriched-prompt content profile keywords)
        response (.generateContent agent prompt)]
    (parse-response response)))

Real Metrics:

• Average time: 5.2s (3.1s LLM + 2.1s queries)
• Average cost: $0.078
• Success rate: 94.7%
• Cache critical: Reduces queries by 85% with 1h TTL

🔷 Type C: AI + Web (Grounding)

Definition

Input + Web Search/Grounding → LLM → Validated output

Agent performs real-time external searches (web, scientific databases, public APIs) to enrich context and validate information. Uses Vertex AI Grounding API or custom tools.

Characteristics

• ✅ External updated data: Doesn't depend on model knowledge
• ✅ Real-time validation: Search evidence, prices, availability
• ✅ Trusted sources: PubMed, Google Scholar, specialized APIs
• ⚠️ High latency: +8-12s for complete grounding
• ⚠️ High cost: $0.15-0.30 per execution
• ⚠️ Rate limits: External APIs may limit requests

When to Use

1. Scientific Research: Search papers, studies, guidelines
2. Fact Validation: Verify claims against authoritative sources
3. Prices/Availability: Query real-time information
4. External Compliance: Check published regulations
5. Public Data Enrichment: Government APIs, datasets

Healthcare Example: S.2-1.2 - Scientific Reference Search

Function: Find scientific references to validate medical claims

Implementation Pattern (Clojure):

(ns lab.agents.scientific-search
  "Type C: AI + Web Grounding - Scientific reference search"
  (:require [lab.config.google-cloud :as gc])
  (:import [com.google.cloud.vertexai.api Tool GroundingMetadata]))

(defn create-grounded-agent
  "Creates Type C agent with grounding enabled."
  [{:keys [project-id location grounding-sources]
    :or {grounding-sources ["google_search" "google_scholar"]}}]
  (let [vertex-ai (VertexAI. project-id location)
        grounding-tool (-> (Tool/newBuilder)
                           (.setGoogleSearchRetrieval
                             (-> (GoogleSearchRetrieval/newBuilder)
                                 (.build)))
                           (.build))
        model-instance (-> (.getGenerativeModel vertex-ai "gemini-1.5-pro")
                           (.withTools [grounding-tool]))]
    model-instance))

(defn search-scientific-references
  "Searches scientific references with grounding."
  [agent claims]
  (let [prompt (format
                 "For each medical claim, search reliable scientific references.
                 Minimum 2 primary sources, 1 systematic review if available.
                 Sources: PubMed, Google Scholar, SciELO. Return JSON."
                 (json/generate-string claims))
        response (.generateContent agent prompt)
        text (-> response .getText)
        grounding-metadata (.getGroundingMetadata response)]
    (assoc (json/parse-string text true)
           :grounding-sources (-> grounding-metadata .getSearchQueries vec)
           :grounding-confidence (-> grounding-metadata .getConfidence))))

Real Metrics:

• Average time: 12.1s (8.2s grounding + 3.9s LLM)
• Average cost: $0.175
• Success rate: 91.7%
• References/claim: 4.3 average
• Important: Rate limiting on public APIs (PubMed: 3 req/s)

🔶 Type D: AI + CAG + Web (Maximum Context)

Definition

Input + Database Context + Web Grounding → LLM → Consolidated output

Agent combines all capabilities: accesses tenant DB AND performs external searches. Maximum possible context, used for final consolidation or complex decisions.

Characteristics

• ✅ Maximum context: Tenant DB + web + input
• ✅ Complex decisions: Multiple sources of truth
• ✅ Total compliance: Internal rules + external regulations
• ⚠️ Maximum latency: 15-20s total
• ⚠️ Maximum cost: $0.30-0.50 per execution
• ⚠️ High complexity: Orchestration of multiple sources

When to Use

1. Final Consolidation: Aggregate data from multiple sources
2. Critical Decisions: Require complete context
3. Multi-Source Validation: Check against DB + web
4. Executive Reports: Insights from internal + external data
5. Total Compliance: Validate against internal rules + public regulations

Healthcare Example: S.4-1.1-3 - Final Text Consolidation

Function: Consolidate all previous outputs (LGPD data, validated claims, references, SEO) into final professional text ready for publication

Real Metrics:

• Average time: 16.8s (4.2s queries + 9.5s grounding + 3.1s LLM)
• Average cost: $0.42
• Success rate: 89.3%
• Trade-off: Maximum quality vs cost/latency

📘 Complete implementation: See cva-healthcare-pipeline skill, System S.4.

📊 Selection Strategy (Decision Tree)

┌─────────────────────────────────────────┐
│ Need external data (web/APIs)?         │
└─────────────┬───────────────────────────┘
              │
        ┌─────┴─────┐
        │           │
       NO          YES
        │           │
        v           v
┌───────────┐  ┌─────────────────────────────┐
│ Need      │  │ Need tenant DB also?        │
│ tenant DB?│  └──────────┬──────────────────┘
└─────┬─────┘             │
      │             ┌─────┴─────┐
  ┌───┴───┐         │           │
  │       │        NO          YES
 NO      YES        │           │
  │       │         v           v
  v       v    ┌────────┐  ┌────────┐
┌───┐  ┌───┐  │ TYPE C │  │ TYPE D │
│ A │  │ B │  │        │  │        │
└───┘  └───┘  └────────┘  └────────┘

Mapping Examples

💡 Best Practices

1. Always Start with Type A

"Premature optimization is the root of all evil" - Donald Knuth

Implement first as Type A. If output is unsatisfactory, consider B/C/D. Often, a well-structured Type A prompt is sufficient.

2. Context Caching (Type B)

DB contexts rarely change. Aggressive caching:

• Professional profiles: TTL 1h
• Regulations: TTL 24h
• SEO keywords: TTL 7 days

Observed savings: 85% query reduction

3. Batch for Type C

Grounding is expensive. Group multiple claims in a single call:

;; ❌ Inefficient: 5 calls
(doseq [claim claims]
  (search-references agent [claim]))

;; ✅ Efficient: 1 call
(search-references agent claims)

Savings: ~70% cost and latency

4. Conditional Grounding (Type D)

Not always need web grounding. Use conditional logic:

(defn consolidate-content
  [agent inputs]
  (let [db-contexts (fetch-all-contexts inputs)
        needs-grounding? (requires-external-validation? inputs)
        web-data (when needs-grounding?
                   (perform-grounding inputs))]
    (generate-final-content agent db-contexts web-data)))

5. Fallback Strategy

Grounding can fail (rate limits, timeout). Have fallback to Type B:

(defn robust-consolidation
  [agent inputs]
  (try
    (consolidate-tipo-d agent inputs)  ; Try with grounding
    (catch Exception e
      (log/warn "Grounding failed, falling back to Type B" e)
      (consolidate-tipo-b agent inputs))))  ; Fallback without web

🎯 Executive Summary

Recommended Distribution in Pipeline (Healthcare):

• Type A: 40% of systems (extraction, classification)
• Type B: 40% (personalization, compliance)
• Type C: 15% (scientific validation)
• Type D: 5% (final consolidation)

ROI Multi-Model + Typing:

• Savings vs "Type D everywhere": 67%
• Savings vs "Claude-only": 41%
• Average pipeline time: 34.2s (vs 87s all Type D)

🔗 Related Skills

• cva-healthcare-pipeline - Practical application of all 4 types ⭐
• cva-patterns-cost - Cost optimization strategies
• cva-patterns-context - Context management (CAG) implementation
• cva-patterns-workflows - Multi-agent orchestration

📘 Additional Documentation

For detailed examples by type, see:

• Type A Examples - Pure AI implementations
• Type B Examples - CAG patterns
• Type C Examples - Grounding strategies
• Type D Examples - Maximum context consolidation
• Decision Tree - Visual selection guide

This taxonomy is validated in production with proven ROI. Use it to make systematic architecture decisions and optimize costs.