Module 5 — Agent Architecture
Agent frameworks, MCP protocol, control flow, and resilience — designing robust AI agent systems.
Topics
What's inside this module
- Agentic AI systems learning path
A concise map for agents: loop, tools, memory, planning, approval, reliability, trust infrastructure, routing, and coordination.
- AI Disruption Beyond Horseless Carriage Thinking
Seven AI-native shifts from better old tools toward operators, agent-accessible environments, skills, services, adaptive learning, and compound systems.
- Agent frameworks
LangGraph, CrewAI — multi-agent orchestration.
- MCP protocol
Model Context Protocol — building tool servers for AI.
- Agent loop & control flow
Observe, decide, act, inspect results, and ask humans at the right boundary.
- Agent tool use as a controlled action cycle
How tools give agents hands and senses, and why each call needs scope, interpretation, and approval boundaries.
- Human Agency Is Not a Button
Design authority boundaries for responsible agent systems, not just approval prompts.
- Smart Is Not Reliable
Why agent reliability comes from reality checks, bounded retries, recovery paths, and human escalation.
- Trust Is Earned Through Evidence, Not Vibes
How agent systems earn trust through scoped tasks, permissions, observability, evaluations, fallbacks, and verification.
- Model Choice Is System Design
How agent systems route tasks by difficulty, risk, cost, latency, privacy, verification needs, and human authority.
- LLM-to-LLM Interaction Patterns
How same-provider, cross-provider, and two-model loops work when roles, turns, evidence, and synthesis are bounded.
- More Agents, More Coordination
Why multi-agent systems need task contracts, evidence checks, conflict handling, synthesis, and human escalation.