Agentic AI Reference Architecture

Model Adaptation · Software component

Centralized-Training Decentralized-Execution Learner

Software componentModel AdaptationModelsarc:CTDEMultiAgentLearner

A multi-agent policy learner that trains with a centralized view of global state and all agents' actions, then extracts individual policies that act on local observations only.

Responsibility. Learns coordinated policies centrally for decentralized execution.

Also known as: CTDE, Multi-agent actor-critic, Centralized training with decentralized execution

Variant of Multi-Agent Policy Learner abstract

When to choose. Choose when training can occur with full observability (e.g., simulation) but deployment requires distributed execution with limited communication.

specializestrainsalternative toMulti-Agent Policy Learner: specializesMulti-Agent Policy LearnerValue Network: trainsValue NetworkIndependent Multi-Agent Learner: alternative toIndependent Multi-Agent …
Direct neighbourhood (hover for relationship types)

Relationships

trains lifecycle

alternative to variability

Classification

Patterns
Centralized critic, decentralized actorsQMIX monotonic value factorization (mixing network)
Quality attributes
Functional suitability: correctness and validity (ISO/IEC 25010 | NIST AI RMF: valid)Performance efficiency (ISO/IEC 25010)
Risks mitigated
Non-stationarity of independent learnersMulti-agent credit assignment

Sources

  1. Ch5.12: T. Nguyen, "Learning-Based Decision Making Fundamentals," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 5.12. ISBN: 9798244538229.
  2. Ch5.13: T. Nguyen, "Hybrid Decision Systems Fundamentals," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 5.13. ISBN: 9798244538229.