LN: Mavor-Parker (2024) — Agentic AI

Bibliographic Reference

Citation: Mavor-Parker, A. (2024). Agentic AI: The next generation of intelligent systems. Manning Publications. https://www.manning.com/books/agentic-ai


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategoryConceptual + practical book
ContextManning’s “In Action” series for the agentic AI era. Covers the conceptual shift from completion-based to goal-directed LLM systems, and the frameworks enabling cyclical autonomous workflows
CorrectnessGrounded in 2023–2024 empirical results; code examples use LangGraph’s stable API
Contributions(1) Formal definition of agentic vs. non-agentic LLM systems; (2) LangGraph for cyclical (non-linear) workflows; (3) Bounded Autonomy framework; (4) Tool integration taxonomy; (5) Human-agent handoff patterns
ClarityVery good — strong conceptual framing with practical implementations

Relevance: ⭐⭐⭐⭐

Provides the conceptual framework for PUMA’s HITL and Bounded Autonomy design. The LangGraph cyclical workflow model informs PUMA Stage 5 (SmartPMO) architecture.


Pass 2 — Key Concepts

Agentic vs. Non-Agentic LLM Systems

PropertyNon-Agentic (Completion)Agentic
InputSingle promptGoals + context
OutputSingle responseActions in environment
MemoryNone (stateless)Persistent (episodic, semantic)
PlanningNoneMulti-step planning
ToolsNoneExternal API calls, file I/O
FeedbackNoneObservation → re-plan

PUMA Stages 1–3 are non-agentic (single-pass LLM completions). Stage 4–5 are agentic (ReAct loop, multi-tool, persistent state).

LangGraph: Cyclical Workflows

Traditional LLM pipelines are directed acyclic graphs (DAGs) — each step runs once. LangGraph enables cyclical graphs: steps can loop back based on conditions.

┌─────────────────────────────────────────────────────┐
│ Issue Input → Triage Agent                          │
│        ↓                                             │
│ Confidence ≥ 0.80? ──Yes──→ Output                  │
│        ↓                                             │
│        No → Reflexion Agent → Revised Triage        │
│        ↓                                             │
│ Confidence ≥ 0.80? ──Yes──→ Output                  │
│        ↓                                             │
│        No → HITL Escalation → Human → Output        │
└─────────────────────────────────────────────────────┘

This is exactly PUMA’s Stage 4+ architecture: a cyclical workflow with confidence-based routing and HITL fallback.

Bounded Autonomy Framework

Mavor-Parker formalises the gradient from full human control to full automation:

LevelDescriptionHuman Role
1Human executes, AI suggestsAll decisions
2AI executes, human approves each stepGatekeeper
3AI executes, human approves batchesAuditor
4AI executes autonomously, human reviews exceptionsException handler
5AI executes fully autonomouslyNone

PUMA targets Level 4 for high-confidence cases and falls back to Level 2–3 for low-confidence issues. This is HITL as an exception-handling pattern, not a bottleneck.

Human-Agent Handoff Patterns

The book identifies four handoff patterns:

  1. Always-on: Human reviews every output (expensive, defensive)
  2. Sampling audit: Human reviews a random sample (statistically controlled)
  3. Threshold-based: Human reviews outputs below a confidence threshold (PUMA’s model)
  4. Drift detection: Human reviews when model performance diverges from baseline

PUMA Integration

  • PUMA Stage 4+: LangGraph cyclical workflow as the implementation target
  • HITL design: Bounded Autonomy Level 4 as the PUMA operational mode
  • Threshold-based escalation: Confidence < 0.80 → HITL (from PN-HITL-BoundedAutonomy)

MOCs