Agentic Design Patterns — Architecture Alignment¶
Full coverage matrix showing alignment between Agentic Design Patterns (Antonio Gulli, 2025) and the
agent-utilitiesconcept architecture.
Coverage Matrix¶
| # | Design Pattern | agent-utilities Concept(s) | Status |
|---|---|---|---|
| 1 | Prompt Chaining | CONCEPT:AU-ORCH.planning.recursion-nesting-depth (Prompt Chain Executor) | ✅ |
| 2 | Routing | CONCEPT:AU-ORCH.execution.inject-signal-board-observations (Graph Orchestration / Router) | ✅ |
| 3 | Parallelization | CONCEPT:AU-ORCH.execution.inject-signal-board-observations (parallel_batch_processor, fan-out/fan-in) | ✅ |
| 4 | Reflection | CONCEPT:AU-ORCH.planning.recursion-nesting-depth (AHE critique), CONCEPT:AU-KG.memory.tiered-memory-caching (Self-Model) | ✅ |
| 5 | Tool Use | CONCEPT:AU-ECO.messaging.native-backend-abstraction (Tool/MCP Registry) | ✅ |
| 6 | Planning | CONCEPT:AU-AHE.harness.harness-evolution (SDD pipeline), Planner node | ✅ |
| 7 | Multi-Agent | CONCEPT:AU-KG.query.object-graph-mapper (Swarm), Council, Teams | ✅ |
| 8 | Memory Management | KG engine, MAGMA views, episodic memory | ✅ |
| 9 | Learning & Adaptation | CONCEPT:AU-ORCH.execution.inject-signal-board-observations (Evolutionary Variants), CONCEPT:AU-ORCH.planning.recursion-nesting-depth (AHE) | ✅ |
| 10 | MCP | CONCEPT:AU-ECO.messaging.native-backend-abstraction, mcp/ module |
✅ |
| 11 | Goal Setting | CONCEPT:AU-AHE.harness.harness-evolution (SDD spec/verify), Verifier node | ✅ |
| 12 | Exception Handling | CONCEPT:AU-OS.state.cognitive-scheduler-preemption (stuck-loop, circuit-breaker) | ✅ |
| 13 | Human-in-the-Loop | ApprovalManager, tool_guard, elicitation | ✅ |
| 14 | Knowledge Retrieval | HybridRetriever, KB layer, MAGMA | ✅ |
| 15 | Inter-Agent Comm | CONCEPT:AU-OS.safety.doom-loop-detection (A2A adapter) | ✅ |
| 16 | Resource Optimization | CONCEPT:AU-OS.state.cognitive-scheduler-preemption (ResourceOptimizer) | ✅ |
| 17 | Reasoning | CONCEPT:AU-ORCH.planning.recursion-nesting-depth (RLM), Council deliberation | ✅ |
| 18 | Guardrails/Safety | security/guardrails.py, tool_guard |
✅ |
| 19 | Evaluation & Monitoring | CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort (EvaluationMonitor) | ✅ |
| 20 | Prioritization | CONCEPT:AU-ORCH.planning.recursion-nesting-depth (PrioritizationEngine) | ✅ |
| 21 | Exploration & Discovery | CONCEPT:AU-AHE.harness.evolutionary-aggregation (ExplorationEngine) | ✅ |
New Concepts (CONCEPT:AU-ORCH.planning.recursion-nesting-depth through CONCEPT:AU-AHE.harness.evolutionary-aggregation)¶
CONCEPT:AU-ORCH.planning.recursion-nesting-depth: Prompt Chaining Pattern¶
Module: agent_utilities/patterns/prompt_chain.py
Declarative multi-step prompt pipelines with intermediate validation, conditional branching, and KG persistence. Each chain step transforms its predecessor's output and optionally validates against a Pydantic model.
Key Classes:
- PromptChainStep — Single step definition with template and branch logic
- PromptChain — Ordered collection of steps with retry configuration
- PromptChainExecutor — Async executor managing state flow between steps
OWL: :PromptChain rdfs:subClassOf :Procedure
CONCEPT:AU-OS.state.cognitive-scheduler-preemption: Resource-Aware Optimization¶
Module: agent_utilities/core/resource_optimizer.py
Cost-aware model selection, per-specialist budget allocation, and latency routing with configurable industry-standard defaults.
Key Classes:
- ResourceBudget — Per-session budget with token/cost/latency tracking
- ResourceOptimizer — Budget allocation, model selection, usage recording
Configuration (env vars):
- SESSION_TOKEN_BUDGET — Default: 500,000 tokens
- SESSION_COST_BUDGET_USD — Default: $5.00
- SESSION_LATENCY_BUDGET_MS — Default: 30,000ms
OWL: :ResourceUsage rdfs:subClassOf bfo:Process
CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort: Evaluation & Monitoring¶
Module: agent_utilities/harness/continuous_evaluation_engine.py
Multi-dimensional evaluation with LLM-as-Judge rubrics, trend monitoring, and quality alerting. The weighted composite score is the current verifier-gate currency.
Key Classes:
- EvaluationDimension — Single dimension score with rubric and evidence
- MultiDimensionalEvaluation — Aggregated dimensions with composite
- EvaluationMonitor — Trend tracking, alerting, KG persistence
Dimensions: correctness (35%), completeness (25%), relevance (25%), safety (15%)
OWL: :EvaluationRecord rdfs:subClassOf :Observation
CONCEPT:AU-ORCH.planning.recursion-nesting-depth: Task Prioritization¶
Module: agent_utilities/patterns/prioritization.py
Multi-factor priority scoring with dependency tracking, priority inheritance, and capability-based specialist assignment.
Key Classes:
- PriorityScore — Multi-factor score (urgency × impact × effort × risk)
- PrioritizationEngine — Task management with reprioritization and ordering
Weight Formula (global defaults): - Urgency: 35%, Impact: 30%, Effort: 20% (inverse), Risk: 15%
OWL: :PrioritizedTask rdfs:subClassOf :Action, :blocks owl:TransitiveProperty
CONCEPT:AU-AHE.harness.evolutionary-aggregation: Exploration & Discovery¶
Module: agent_utilities/patterns/exploration.py
Autonomous exploration loop for arbitrary domain discovery (code, medical, finance, etc.) with hypothesis generation, experiment design, and multi-reviewer evaluation.
Key Classes:
- KnowledgeGap — Identified knowledge deficit with severity scoring
- Hypothesis — Testable prediction linked to a gap
- Experiment — Structured test with variables and success criteria
- ExplorationEngine — Full lifecycle management from gap → discovery
Lineage: KnowledgeGap → Hypothesis → Experiment → Discovery → KBFact
OWL: :Experiment rdfs:subClassOf :Procedure, :testsHypothesis, :exploredGap
KG Node & Edge Types Added¶
Node Types¶
| Type | Concept | Description |
|---|---|---|
PROMPT_CHAIN |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Prompt pipeline definition with steps |
RESOURCE_USAGE |
CONCEPT:AU-OS.state.cognitive-scheduler-preemption | Per-session resource consumption record |
EVALUATION_RECORD |
CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort | Multi-dimensional evaluation result |
PRIORITIZED_TASK |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Task with multi-factor priority scoring |
KNOWLEDGE_GAP |
CONCEPT:AU-AHE.harness.evolutionary-aggregation | Identified knowledge deficit |
EXPLORATION_EXPERIMENT |
CONCEPT:AU-AHE.harness.evolutionary-aggregation | Experiment testing a hypothesis |
Edge Types¶
| Type | Concept | Description |
|---|---|---|
CHAIN_STEP |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Ordered step within a prompt chain |
BRANCHES_TO |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Conditional branch between steps |
CONSUMED_RESOURCE |
CONCEPT:AU-OS.state.cognitive-scheduler-preemption | Episode → resource usage link |
EVALUATED_WITH |
CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort | Episode → evaluation record link |
CALIBRATED_AGAINST |
CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort | Machine vs human evaluation link |
BLOCKS |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Task blocking relationship |
ASSIGNED_TO_SPECIALIST |
CONCEPT:AU-ORCH.planning.recursion-nesting-depth | Task → specialist assignment |
TESTS_HYPOTHESIS |
CONCEPT:AU-AHE.harness.evolutionary-aggregation | Experiment → hypothesis link |
EXPLORED_GAP |
CONCEPT:AU-AHE.harness.evolutionary-aggregation | Hypothesis → knowledge gap link |
RESULTED_IN_DISCOVERY |
CONCEPT:AU-AHE.harness.evolutionary-aggregation | Experiment → fact/discovery link |