📘 Demystifying AI Agents – Microsoft Training (Cihan Cinar, 5/21/2025)

​

Overview
This session unpacked what AI agents are, how they differ from simple models, and how Microsoft tools support building and deploying them. Cihan Cinar (Software & AI Architect) emphasized both the architecture of agents and the practical risks and opportunities they bring.

​

Key Concepts

• Definition: AI Agents are systems that perceive inputs, make decisions, take actions, and adapt through feedback.

• Core Components:

  • Perception (data, sensors, user input)

  • Reasoning (rules, LLMs, planning, ML models)

  • Action (executing tasks, tool/API integration)

  • Learning (adapting via reinforcement or supervised learning)

• Agent Architectures: modular design with pipelines, memory, and orchestration. Multi-agent systems allow agents to collaborate.

• Microsoft Integration: Azure AI Agent Service and orchestration frameworks for building scalable, tool-using, and governed agents.

​

Challenges & Risks

• Hallucinations and ambiguity in LLM-based agents.

• Trust, transparency, and the need for observability/logging.

• Security, privacy, and bias concerns.

• Importance of clear scope and responsible autonomy.

​

Best Practices

• Define the agent’s purpose clearly and limit scope.

• Build a minimal viable agent first, then expand capabilities.

• Include monitoring, feedback, and human-in-the-loop safeguards.

• Use sensitivity labeling, policy enforcement, and compliance frameworks for responsible deployment.