AI Agents & LLM Apps: Agent Loops, Tools/MCP, Agentic RAG, Security, Evals & Troubleshooting
The 2026 "AI system design" round: build agents as distributed systems with budgets, validated tools, durable steps, least privilege and evals, then debug them when they are slow, wrong, stuck or expensive.
About 180 minutes · 20 exam questions · 14 interview questions · 2 code labs
What you will learn
- The agent loop: plan, act, observe, and the budgets that make it stop
- Tool calling and tool schemas: designing the agent-computer interface
- MCP: servers, transports and auth, and tool discovery when you have hundreds of tools
- Context engineering: prefix caching economics, window budgeting, and long context vs RAG
- Memory: short-term transcript, compaction, and long-term memory that does not rot or get poisoned
- Agentic RAG vs classic RAG: when retrieval becomes a tool the model drives
- Multi-agent orchestration, and when NOT to use it
- Durable agent execution: checkpoints, idempotent tools and human-in-the-loop approvals
- Security: prompt injection, least-privilege tools, exfiltration and sandboxing
- Cost and latency control: model routing, caching, streaming and parallel tool calls
- Evals as a system: tool-call, trajectory and outcome evals, LLM judges, regression gates, online metrics
- Observability: tracing agent runs step by step
- Worked design: a customer-support agent for an Indian fintech (requirements, estimation, architecture, deep dives)
- Troubleshooting playbook: slow chatbot, wrong RAG answers, looping agents, tripled costs
Previous: Instagram & Food Delivery (Swiggy / Zomato / DoorDash)