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)
| C | Assessment |
|---|---|
| Category | Conceptual + practical book |
| Context | Manning’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 |
| Correctness | Grounded 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 |
| Clarity | Very 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
| Property | Non-Agentic (Completion) | Agentic |
|---|---|---|
| Input | Single prompt | Goals + context |
| Output | Single response | Actions in environment |
| Memory | None (stateless) | Persistent (episodic, semantic) |
| Planning | None | Multi-step planning |
| Tools | None | External API calls, file I/O |
| Feedback | None | Observation → 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:
| Level | Description | Human Role |
|---|---|---|
| 1 | Human executes, AI suggests | All decisions |
| 2 | AI executes, human approves each step | Gatekeeper |
| 3 | AI executes, human approves batches | Auditor |
| 4 | AI executes autonomously, human reviews exceptions | Exception handler |
| 5 | AI executes fully autonomously | None |
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:
- Always-on: Human reviews every output (expensive, defensive)
- Sampling audit: Human reviews a random sample (statistically controlled)
- Threshold-based: Human reviews outputs below a confidence threshold (PUMA’s model)
- 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)
Related Notes
- LN-Shieh-2024-GenerativeAIAgents — implementation complement: Shieh for code, Mavor-Parker for architecture
- PN-HITL-BoundedAutonomy — Bounded Autonomy permanent note
- LN-Wooldridge-2009-MultiAgentSystems — theoretical MAS foundations
- Smart-PMO-Vision — SmartPMO as a Level 4 Bounded Autonomy system